Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
i
FILIPE ANDRÉ DE SOUSA FIGUEIRA BARATA
Licenciado
Control in Distribution Networks with Demand Side Management
Dissertação para obtenção do Grau de Doutor em Engenharia Electrotécnica e de Computadores
Orientador: Prof. Doutor Rui Alexandre Nunes Neves da Silva, Professor Auxiliar, Faculdade de Ciências e Tecnologias da Universidade Nova de Lisboa
Júri:
Presidente: Prof. Doutor Jorge Joaquim Pamies Teixeira
Arguente(s): Prof. Doutora Teresa Maria de Gouveia Torres Feio Mendonça
Prof. Doutor João Francisco Alves Martins
Vogais: Prof. Doutor José Manuel Prista do Valle Cardoso Igreja
Prof. Doutor Miguel José Simões Barão
Dezembro, 2015
i
Control in Distribution Networks with Demand Side Management
Copyright © Filipe André Barata, Faculdade de Ciências e Tecnologia, Universidade Nova
de Lisboa.
The Faculdade de Ciências e Tecnologia and the Universidade Nova de Lisboa have the
perpetual right, without geographical limits, to archive and publish this dissertation either in
print or digital form, or any other medium that is still to be invented, and distribute it through
scientific repositories, admitting its copy and distribution for educational or research purposes,
as well as for non-commercial purposes, as long as the author and the editor are credited for
their work.
A Faculdade de Ciências e Tecnologia e a Universidade Nova de Lisboa têm o direito,
perpétuo e sem limites geográficos, de arquivar e publicar esta dissertação através de
exemplares impressos reproduzidos em papel ou de forma digital, ou por qualquer outro meio
conhecido ou que venha a ser inventado, e de a divulgar através de repositórios científicos e de
admitir a sua cópia e distribuição com objetivos educacionais ou de investigação, não
comerciais, desde que seja dado crédito ao autor e editor.
ii
iii
Acknowledgements
I wish to express my gratitude to my Supervisor Professor Rui Neves-Silva for the opportunity
he gave me and for the confidence he deposited in me, accepting this project inspired by him. I
also like to thank his support and understanding in the most difficult periods of this journey, and
by all his collaboration and advices which guided me until this moment.
Similarly, I would like to reserve a special acknowledgment to Professor José Manuel Igreja, for
his priceless academic, professional, personal guidance and encouragement.
For this work I would like also to thank my CAT committee member Dr. João Martins, for is
time, interest, and helpful comments.
I wish to personally thank my colleagues and friends from ISEL: Carla Viveiros, José Ribeiro,
Luís Encarnação, Nuno Domingues, Nuno Guedes, Pedro Fonte, Ricardo Luís e Rita Pereira.
These are fantastic people that provide me priceless moments and with whom I have the
privilege to share a considerable part of my days since I have started my academic life.
Finally, and most importantly, I am thankful for the emotional support of my family. To Manuel
Abreu Rodrigues, a tremendous friend that recently passed away, all my gratitude for his
contagious joy that had provided me and my family unforgettable moments. His unexpected
passing left me emotionally wounded, but allowed me to rethink better about my life. He will be
sorely missed. To my Mother and Grandmother; two super women, an example of strength and
determination. Despite all the drawbacks life brought them, they rose up and continued.
To my wife Paula, to my son Salvador and to the newly arrived daughter Joana, my gratitude for
being there every day with a smile and for the comfort and encouragement words, and also to
them, my apologies for all the absences required by this work.
iv
v
Abstract
The way in which electricity networks operate is going through a period of significant change.
Renewable generation technologies are having a growing presence and increasing penetrations
of generation that are being connected at distribution level. Unfortunately, a renewable energy
source is most of the time intermittent and needs to be forecasted.
Current trends in Smart grids foresee the accommodation of a variety of distributed generation
sources including intermittent renewable sources. It is also expected that smart grids will
include demand management resources, widespread communications and control technologies
required to use demand response are needed to help the maintenance in supply-demand balance
in electricity systems. Consequently, smart household appliances with controllable loads will be
likely a common presence in our homes. Thus, new control techniques are requested to manage
the loads and achieve all the potential energy present in intermittent energy sources.
This thesis is focused on the development of a demand side management control method in a
distributed network, aiming the creation of greater flexibility in demand and better ease the
integration of renewable technologies. In particular, this work presents a novel multi-agent
model-based predictive control method to manage distributed energy systems from the demand
side, in presence of limited energy sources with fluctuating output and with energy storage in
house-hold or car batteries. Specifically, here is presented a solution for thermal comfort which
manages a limited shared energy resource via a demand side management perspective, using an
integrated approach which also involves a power price auction and an appliance loads allocation
scheme.
The control is applied individually to a set of Thermal Control Areas, demand units, where the
objective is to minimize the energy usage and not exceed the limited and shared energy
resource, while simultaneously indoor temperatures are maintained within a comfort frame.
Thermal Control Areas are overall thermodynamically connected in the distributed environment
and also coupled by energy related constraints. The energy split is performed based on a fixed
sequential order established from a previous completed auction wherein the bids are made by
each Thermal Control Area, acting as demand side management agents, based on the daily
energy price. The developed solutions are explained with algorithms and are applied to different
scenarios, being the results explanatory of the benefits of the proposed approaches.
Keywords: DMPC, MAS, intermittent energy resource, DSM, energy auction, thermal control
areas, shifting loads, energy efficiency;
vi
vii
Resumo
A forma como as redes elétricas operam está a atravessar um período de transformação
significativa. As tecnologias de geração renováveis possuem uma presença crescente e a
penetração desta geração ao nível da distribuição está a aumentar. Infelizmente, uma fonte de
energia renovável é na maioria do tempo intermitente e necessita de algum tipo de previsão e
antecipação de comportamento.
As têndencia actuais redes elétricas inteligentes preveêm que estas irão acomodar de uma forma
integrada, formas de armazenamento de energia e uma variedade de fontes de geração
distribuídas incluindo fontes renováveis intermitentes. É também expectável que as redes
inteligentes irão incluir sistemas de gestão da procura e a difusão da comunicações e tecnologias
de controlo necessários para que a resposta à procura auxilie no equilíbrio entre oferta e
demanda em sistemas de energia elétrica. Eletrodomésticos inteligentes serão provavelmente
uma presença comum nos nossos lares.
Desta forma, são necessárias novas técnicas de controlo para gerir as cargas por forma a
aproveitar todo o potencial energético existente em fontes de energia intermitentes.
Esta tese foca uma metodologia de controlo para gestão da procura em redes distribuídas, para
criar maior flexibilidade do lado da procura e facilitar a integração de tecnologias renováveis.
Em particular, o trabalho apresenta um novo método de controlo predictivo multi-agente para
gestão sistemas de redes de energia distribuídas do lado da procura, quando na presença de
recursos energéticos limitados e flutuantes, e com armazenamento de energia em baterias
domésticas ou de veículos. Especificamente, é aqui apresentada uma solução para conforto
térmico que gere um recurso energético limitado e partilhado numa perspetiva de gestão da
procura, utilizando uma abordagem integrada que envolve um leilão de preço de energia e um
esquema de alocação de cargas domésticas.
O controlo é aplicado individualmente a um conjunto de Áreas de Controlo Térmico, unidades
de demanda (consumidores), onde o objetivo é minimizar a utilização de energia sem exceder o
recurso partilhado e limitado enquanto, simultaneamente, a temperatura interior é mantida
dentro da zona de conforto. Em ambiente distribuído, as Áreas de Controlo Térmico estão
geralmente termodinamicamente ligadas e também acopladas pela restrição energética. A
distribuição da energia é efetuada com base numa ordem sequencial fixa estabelecida por um
leilão onde cada uma das Áreas de Controlo Térmico, atuando como agentes de gestão da
procura, faz as suas licitações com base no preço diário da energia. A solução desenvolvida é
viii
explicada por algoritmos e aplicada a diferentes cenários onde os resultados obtidos ilustram os
benefícios da abordagem proposta.
Palavras chave: Controlo Predictivo Distribuído, Sistemas Multi-Agente, recurso de energia
intermitente, gestão da procura, leilão de energia, áreas de controlo térmico, alocação de cargas,
eficiência energética;
ix
Table of Contents
Acknowledgements ............................................................................................. iii
Abstract ................................................................................................................ v
Resumo ............................................................................................................... vii
Table of Contents ............................................................................................... ix
List of Figures ................................................................................................... xiii
List of Tables ..................................................................................................... xix
List of Acronyms and Abbreviations ............................................................. xxi
1 Introduction .................................................................................................. 1
1.1 Introduction .................................................................................................................. 1
1.2 Background .................................................................................................................. 2
1.3 Motivation .................................................................................................................... 3
1.4 Research Questions ...................................................................................................... 5
1.5 Aimed Contributions.................................................................................................... 5
1.6 Outline ......................................................................................................................... 7
2 Literature Review and Model Based Predictive Control ....................... 11
2.1 Introduction ................................................................................................................ 11
2.2 Demand Side Management for Distribution Networks.............................................. 13
2.3 Model Based Predictive Control ................................................................................ 18
2.3.1 Linear Plant Model .......................................................................................................... 22
2.3.2 Nonlinear Plant Model .................................................................................................... 28
2.3.3 Generalized Predictive Control ....................................................................................... 29
2.3.4 Stability and Feasibility ................................................................................................... 33
2.3.5 Centralized MPC ............................................................................................................. 35
x
2.3.6 Decentralized MPC ......................................................................................................... 37
2.3.7 Distributed Model Predictive Control ............................................................................. 39
2.4 Multi-Agents systems (MAS) .................................................................................... 44
2.4.1 MAS architectures ........................................................................................................... 45
3 Dynamical Models and Scenarios Description ........................................ 51
3.1 Introduction ................................................................................................................ 51
3.2 Scenarios overview .................................................................................................... 52
3.3 Demand side management approach ......................................................................... 56
3.4 TCA Dynamical Models ............................................................................................ 57
4 MPC and PI control in thermal comfort systems ................................... 65
4.1 Introduction ................................................................................................................ 65
4.2 System description ..................................................................................................... 66
4.3 Results ........................................................................................................................ 74
4.3.1 PI versus MPC. ................................................................................................................ 76
4.3.2 MPC with “comfort zone” ............................................................................................... 77
4.3.3 MPC with “comfort zone” and constrained available power .......................................... 79
4.3.4 Conclusions ..................................................................................................................... 80
5 Thermal Comfort with Demand Side Management Using
Distributed MPC ............................................................................................... 81
5.1 Introduction ................................................................................................................ 81
5.2 Implemented global scenario ..................................................................................... 82
5.3 MPC formalization .................................................................................................... 84
5.3.1 Algorithm I - Implemented sequential scheme ................................................................ 90
6 DMPC for Thermal House Comfort with Sequential Access
Auction ............................................................................................................... 91
6.1 Introduction ................................................................................................................ 91
6.2 Access order with fixed stablished sequence ............................................................. 92
xi
6.2.1 Results ............................................................................................................................. 93
6.3 Access order with variable hourly sequence ............................................................ 110
6.3.1 Algorithm II - Implemented sequential scheme with variable hourly sequence ........... 111
6.3.2 Results ........................................................................................................................... 112
6.4 Conclusions .............................................................................................................. 119
7 DMPC for Thermal House Comfort with Sliding Load ....................... 121
7.1 Introduction .............................................................................................................. 121
7.2 Algorithm III – Implemented shifting and loads allocation scheme ........................ 123
7.3 Results ...................................................................................................................... 126
7.3.1 One house scenario ........................................................................................................ 127
7.3.2 Distributed scenario ....................................................................................................... 131
7.4 Conclusions .............................................................................................................. 139
8 Conclusions and Future Work Directions ............................................. 141
8.1 Conclusion and Summary of Achievements ............................................................ 141
8.2 Recommendations for Future Work Directions ....................................................... 143
Bibliography .................................................................................................... 145
xii
xiii
List of Figures
Figure 2.1. Smart Grid Roadmap (Bsria, 2014) .......................................................................... 12
Figure 2.2. Venn diagram with implemented technologies. ........................................................ 12
Figure 2.3. Basic block diagram of MPC .................................................................................... 20
Figure 2.4. Conceptual picture of the moving horizon in predictive control (Boom &
Stoorvogel, 2010). .............................................................................................................. 21
Figure 2.5. Hard constraint representation (Adapted from Froisy, 1994). .................................. 25
Figure 2.6. Soft constraint representation (Adapted from Froisy, 1994). .................................... 25
Figure 2.7. GPC control structure. .............................................................................................. 31
Figure 2.8. Disturbance load profile in the several prediction horizons. .................................... 34
Figure 2.9. Indoor temperature profile for the several prediction horizons. ............................... 34
Figure 2.10. Power profile for the several prediction horizons. .................................................. 34
Figure 2.11. Consumption profile for several prediction horizons. ............................................ 35
Figure 2.12. Centralized MPC architecture (Scattolini, 2009). ................................................... 36
Figure 2.13. Decentralized MPC architecture (Scattolini, 2009). ............................................... 38
Figure 2.14. Distributed MPC architecture (Scattolini, 2009). ................................................... 40
Figure 2.15. Single-agent control structure. (Adapted from Negenborn 2007). ......................... 46
Figure 2.16. Multi-agent single layer control structure. (Adapted from Negenborn 2007). ....... 47
Figure 2.17. Multi-layer control structure. (Adapted from Negenborn 2007). ........................... 47
Figure 3.1. Typical smart grid architecture. ................................................................................ 51
Figure 3.2. Typical energy distribution communication architecture for smart grids (Adapted
from Siemens, 2014). ......................................................................................................... 52
Figure 3.3. Implemented scheme. ............................................................................................... 53
Figure 3.4. Thermal Control Area (TCA) smart thermostat controller vision. ........................... 54
Figure 3.5. Thermal Control Area concept.................................................................................. 54
Figure 3.6. Example of a TCA conceptual framework. .............................................................. 55
Figure 3.7. Energy modelling and building design process example with EnergyPlus. ............. 58
xiv
Figure 3.8. Schematic representation of thermal-electrical modular analogy for one division. .. 60
Figure 3.9. Generic schematic representation of thermal-electrical modular analogy for several
divisions (Barata et al., 2014b). .......................................................................................... 60
Figure 3.10. TCA divisions interaction simplified scheme; ........................................................ 62
Figure 3.11. Generalized house/TCA scheme example. ............................................................. 63
Figure 4.1. Block diagram of the implemented system. .............................................................. 67
Figure 4.2. MPC with constraints, system block diagram ........................................................... 72
Figure 4.3. Outdoor temperature. ................................................................................................ 75
Figure 4.4. Indoor temperature. ................................................................................................... 76
Figure 4.5. Power and energy consumption. ............................................................................... 77
Figure 4.6. Temperature error with and without (MPC). ............................................................ 77
Figure 4.7. Case 1 - Indoor temperature with “comfort zone”. ................................................... 78
Figure 4.8. Case 2 - Indoor temperature with “comfort zone”. ................................................... 78
Figure 4.9. Power and energy consumption with “comfort zone”. ............................................. 79
Figure 4.10. Indoor temperature evolution with limited power. ................................................. 79
Figure 4.11. Power and energy consumption with limited power. .............................................. 80
Figure 5.1. Implemented sequential architecture scheme. .......................................................... 83
Figure 5.2. Daily consumer profile. ............................................................................................ 86
Figure 5.3. Thermally coupled TCA’s. ....................................................................................... 86
Figure 6.1: Heat transfer between divisions example. ................................................................ 92
Figure 6.2: System implementation scheme block diagram. ....................................................... 93
Figure 6.3. Outdoor temperature forecasting .............................................................................. 93
Figure 6.4. Scenario I - Disturbance forecasting and indoor temperature A1. ............................ 94
Figure 6.5. Scenario I - Disturbance forecasting and indoor temperature A2. ............................ 95
Figure 6.6. Scenario I - Disturbance forecasting and indoor temperature A3. ............................ 95
Figure 6.7. Scenario I - A1 power profile. ................................................................................... 95
Figure 6.8. Scenario I - A2 power profile. ................................................................................... 96
Figure 6.9. Scenario I - A3 power profile. ................................................................................... 96
xv
Figure 6.10. Scenario I - Global consumption characterization. ................................................. 97
Figure 6.11. Scenario I - Power profile. ...................................................................................... 97
Figure 6.12. Scenario I - Heating/cooling total cost. ................................................................... 97
Figure 6.13. Scenario II - Disturbance forecasting and indoor temperature A1. ......................... 98
Figure 6.14. Scenario II - Disturbance forecasting and indoor temperature A2. ......................... 98
Figure 6.15. Scenario II - Disturbance forecasting and indoor temperature A3. ......................... 99
Figure 6.16. Scenario II – A1 power profile. ............................................................................... 99
Figure 6.17. Scenario II – A2 power profile. ............................................................................... 99
Figure 6.18. Scenario II – A3 power profile. ............................................................................. 100
Figure 6.19. Scenario II - Global consumption characterization............................................... 100
Figure 6.20. Scenario II - Power profile. ................................................................................... 101
Figure 6.21. Scenario II - Heating/cooling total cost. ............................................................... 101
Figure 6.22. Scenario II - Batteries profile. ............................................................................... 101
Figure 6.23. Scenario III - Disturbance forecasting and indoor temperature A1. ...................... 102
Figure 6.24. Scenario III - Disturbance forecasting and indoor temperature A2. ...................... 103
Figure 6.25. Scenario III - Disturbance forecasting and indoor temperature A3. ...................... 103
Figure 6.26. Scenario III – A1 power profile. ............................................................................ 103
Figure 6.27. Scenario III – A2 power profile. ............................................................................ 104
Figure 6.28. Scenario III – A3 power profile. ............................................................................ 104
Figure 6.29. Scenario III - Global consumption characterization. ............................................ 104
Figure 6.30. Scenario III - Power profile. ................................................................................. 105
Figure 6.31. Scenario III - Heating/cooling total cost. .............................................................. 105
Figure 6.32. Scenario IV - Disturbance forecasting and indoor temperature A1. ...................... 106
Figure 6.33. Scenario IV - Disturbance forecasting and indoor temperature A2 ....................... 107
Figure 6.34. Scenario IV - Disturbance forecasting and indoor temperature A3 ....................... 107
Figure 6.35. Scenario IV – A1 power profile. ............................................................................ 107
Figure 6.36. Scenario IV - A2 power profile.............................................................................. 108
Figure 6.37. Scenario IV - A3 power profile.............................................................................. 108
xvi
Figure 6.38. Scenario IV - Global consumption characterization. ............................................ 108
Figure 6.39. Scenario IV - Power profile. ................................................................................. 109
Figure 6.40. Scenario IV - Heating/cooling total cost. .............................................................. 109
Figure 6.41. Daily heating/cooling total cost. ........................................................................... 110
Figure 6.42. Outdoor temperature forecasting (Toa). ................................................................. 113
Figure 6.43. Fixed consumption profile 𝐶𝑤11, and priority level of TCA1. ............................ 113
Figure 6.44. Fixed consumption profile 𝐶𝑤12, and priority level of TCA2. ............................ 113
Figure 6.45. Fixed consumption profile 𝐶𝑤13, and priority level of TCA3. ............................ 114
Figure 6.46. Implemented system. ............................................................................................ 114
Figure 6.47. Access order. ......................................................................................................... 116
Figure 6.48. Disturbance forecasting and indoor temperature A1. ............................................ 116
Figure 6.49. Disturbance forecasting and indoor temperature A2. ............................................ 116
Figure 6.50. Disturbance forecasting and indoor temperature A3. ............................................ 117
Figure 6.51. Consumption and constraints to heat/cool A1. ...................................................... 117
Figure 6.52. Consumption and constraints to heat/cool A2. ...................................................... 118
Figure 6.53. Consumption and constraints to heat/cool A3. ...................................................... 118
Figure 6.54. Global consumption. ............................................................................................. 118
Figure 6.55. Heating/cooling total cost. .................................................................................... 119
Figure 7.1. Shifting load communication infrastructure. .......................................................... 122
Figure 7.2. Implemented shifting load scheme. ........................................................................ 122
Figure 7.3. Total consumption characterization. ....................................................................... 123
Figure 7.4. Implemented power distribution scheme starting in the Optimization Problem 1
(OPi) to OPNS. ................................................................................................................... 124
Figure 7.5.Outdoor temperature forecasting (Toa). .................................................................... 127
Figure 7.6. Thermal disturbance forecasting. ............................................................................ 127
Figure 7.7. Possible loads schedule combinations (PLSCs). ..................................................... 128
Figure 7.8. Total energy costs of FLSCS. .................................................................................. 129
Figure 7.9. Maximum available green energy and chosen sequence ........................................ 130
xvii
Figure 7.10. Indoor temperature. ............................................................................................... 130
Figure 7.11. Used power to heat/cool the space and the maximum green resource available for
comfort. ............................................................................................................................ 131
Figure 7.12. Fixed consumption profile 𝐶𝑤11, and priority level of TCA1. ............................ 132
Figure 7.13. Fixed consumption profile 𝐶𝑤12, and priority level of TCA2. ............................ 133
Figure 7.14. Fixed consumption profile 𝐶𝑤13, and priority level of TCA3. ............................ 133
Figure 7.15. Access order. ......................................................................................................... 133
Figure 7.16. Disturbance forecasting and indoor temperature A1. ............................................ 134
Figure 7.17. Disturbance forecasting and indoor temperature A2. ............................................ 134
Figure 7.18. Disturbance forecasting and indoor temperature A3. ............................................ 135
Figure 7.19. Power profile A1. .................................................................................................. 135
Figure 7.20. Control input profile A1 ........................................................................................ 136
Figure 7.21. Power profile A2. .................................................................................................. 136
Figure 7.22. Control input profile A2. ....................................................................................... 137
Figure 7.23. Power profile A3. .................................................................................................. 137
Figure 7.24. Control input profile A3. ....................................................................................... 138
Figure 7.25. Batteries profile..................................................................................................... 138
Figure 7.26. Consumption costs. ............................................................................................... 138
xviii
xix
List of Tables
Table 2.1. Simulation parameters for the several prediction horizons ........................................ 33
Table 2.2. Computational efforts under several prediction horizons .......................................... 35
Table 4.1.Simulation parameters for PIvsMPC ........................................................................... 76
Table 4.2. Simplified algorithm for MPC with “comfort zone”.................................................. 78
Table 6.1. Distributed parameters ............................................................................................... 94
Table 6.2. Scenario I - Penalty values ......................................................................................... 94
Table 6.3. Scenario III - Penalty values .................................................................................... 102
Table 6.4. Scenario III - Cost comparisons ............................................................................... 105
Table 6.5. Scenario IV - Penalty values .................................................................................... 106
Table 6.6. Example of hourly access order sequence to green energy ...................................... 111
Table 6.7. Bid value for each consumption level by agent. ...................................................... 114
Table 6.8. Scenario parameters. ................................................................................................ 115
Table 7.1.Scenario parameters .................................................................................................. 127
Table 7.2. Shifted loads characteristics ..................................................................................... 128
Table 7.3. Feasible Loads Sequence Combinations .................................................................. 129
Table 7.4. Distributed scenario parameters ............................................................................... 131
Table 7.5. Bid value for each consumption level by TCA ........................................................ 132
Table 7.6. Shifted loads characteristics for distributed scenario ............................................... 132
Table 8.1. Developed algorithm characteristics ........................................................................ 143
xx
xxi
List of Acronyms and Abbreviations
AMI Advanced Metering Infrastructure
ARMAX Autoregressive–Moving-Average Model with exogenous
inputs model
CPP Critical Peak Pricing
DER Distributed Energy Resources
DG Distributed Generation
DMS Distribution Management System
DR Demand Response
DMPC Distributed Model Predictive Control
DSM Demand Side Management
EMS Energy Management System
FLSC Feasible Loads Schedule Combinations
GPC Generalized Predictive Control
HVAC Heating Ventilation and Air Conditioning
I&C Interruptible and Curtailable Load
IL Interruptible Load
LQPC Linear-Quadratic predictive control
LTI Linear Time Invariant
MAS Multi Agent System
MBPC Model Based Predictive Control
MIMO Multiple-Input Multiple-Output
MO Market operator
MPC Model Predictive Control
MUSMAR Multistep Multivariable Adaptive Regulator
OCGT Open Cycle Gas Turbine
PEV Plug-in Electric Vehicle
xxii
PLC Power Line Communication
PLSC Possible Loads Schedule Combinations
PTR Peak Time Rebate
RES Renewable Energy Sources
RHC Receding Horizon Control
RTP Real Time Pricing
SGPC Stable Generalised Predictive Control
SISO Single-Input Single-Output
TCA Thermal Control Area
TCL Thermostatically Controlled Loads
TOU Time Of Use
ZOH Zero-Order Hold
1
Chapter 1
1 Introduction
1.1 Introduction
Traditional electric power systems consist of large power generating plants, interconnected via a
high-voltage transmission lines, loading serving entities which deliver power to end users at
lower voltages using local distribution networks.
Importance of distributed generators (DGs) has increased significantly over the past few years
due to its potential for increasing reliability and lowering the cost of power through the use of
on-site generation. Development of small modular generation technologies, such as
photovoltaic, wind turbines and fuel cells has also contributed to this trend. However, novel
operational and control concepts are needed to make a proper integration into the power grid
system. Control strategies must be further developed to achieve the targeted benefits while
avoiding negative effects on system reliability and safety. The current power distribution system
was not designed to support distribution generation and storage devices at the distribution level;
compatibility, reliability, power quality, system protection and many other issues must also be
considered before the benefits of DGs can be fully obtained (Al-Hinai & Feliachi 2009).
A new emerging type of grid will soon be a reality. Smart Grids (SGs) are planned to include
decentralized generation, active network management for generation and storage, where the
actions of all connected agents to the electricity system can be intelligently integrated aiming for
a sustainable, efficient and secure energy supply system. Future SGs are anticipated to support a
variety of DG, including intermittent renewable sources (James & Jones, 2010).
Smart Grid concept is built on many of the technologies already used by electric utilities but
adds additional communication and control capabilities optimizing the operation of the entire
INTRODUCTION
2
electrical grid. Smart Grid is also positioned to take advantage of new technologies, such as
hybrid plug-in electric vehicles (PEV), various forms of distributed generation, solar energy,
smart metering, lighting management systems, distribution automation and many more.
In a scenario with strong presence of intermittent renewable energy sources (RES) that supply a
set of houses, the new techniques must developed to deal with smart devices, energy
management systems (EMS) such as programmable controllable thermostats and/or PEV
charge, and be capable of making intelligent decisions based on smart prices and users comfort.
Therefore, it is crucial to take into account a global solution that is able to take advantage of
these features.
1.2 Background
Since the beginning of electricity grids, demand has fluctuated and supply has been provided to
follow along. The intermittency of RES, such as wind or solar generation, stills nowadays the
major challenge for the integration of these resources. The present trend in power systems is to
connect more and more RES increasing significantly the security and the quality of supply.
One of the most common solutions to mitigate the problem is operating energy sources, such as
gas turbines or hydro power plants, balancing the variations. But, when the variations are
significant and change rapidly, the backup source needs to be higher and faster, leading to
increased carbon emissions. Also, widely used the frequency regulation is the direct measure of
the balance between generation and system demand at any one instant, and must be maintained
continuously within narrow statutory limits.
Another solution is the energy storage (Ribeiro et al. 2001; Koeppel & Korpas, 2006), if there is
a surplus in energy produced it is fed into the storage. Energy storage is especially critical when
managing the output of intermittent renewable resources, ensuring their generation capacity is
available when needed most, maximising their value, but presently technology available does
only provide answers in low generation time intervals. For example, depending on the size of
the battery pack, electric cars may store enough electricity to power a house for a few hours, and
with small modifications, car batteries can deliver stored power to a home and to the power grid.
As showed there are several solutions to overcome the issue of intermittency but all have
problems to be overcome. The desirable approach is that the demand should follow the source,
passing the control to the demand side, having demand side management (DSM) (Callaway,
2009; Hamidi, 2007). Due its benefits to consumers, enterprises, utilities and society, in the last
decades, many efforts have been made to having demand play a more active role in balancing
the system. DSM benefits are related with customer energy bills and peak power prices
INTRODUCTION
3
decrease, reduction in the need for new power plants, transmission and distribution, reduction in
air pollution and with a significantly increased in energy efficiency. With the growing interest in
intermittent RES use, new opportunities and challenges are triggering and obliging these efforts
to come true. Because of technology barriers and lack of automation, the traditional way of
implementing DSM is via price incentives, i.e. by lowering tariffs at times when the aggregated
demand is expected to be below average, so as to encourage the end user to shift flexible loads
towards these periods (Saffre, 2010). But, with the new technologic advances and world trends
around the SGs concept (Chebbo, 2007; Commission, 2006), and requiring this kind of grid
intelligent control and management based on advanced communication, monitoring solutions,
automation and metering, novel opportunities and challenges to DSM are emerging.
In the smart world, simple household appliances like dishwashers, clothes dryers, simple
electric heaters or heating ventilation or air conditioning (HVAC) systems are planned to be
fully controllable to achieve the network maximum efficiency. Renewable energy sources are
expected to be a common presence and all kWh provided by these technologies should be
efficiently applied. Active demand side management provide solutions to control the loads,
adapting them to the current RES.
Being an actual theme, and despite the present trends and R&D efforts, SGs are still in the
“implementation phase”. More research is needed to provide solid solutions to overcome all the
constraints and turn into reality this ambitious vision. Smart Grids involve a mix of concepts:
distributed generation, DSM, intelligent control, energy efficiency, intermittent RES, thermal
comfort, load control and energy saving are some of them which are interconnected and should
be integrated into solutions. Smart Grids are, therefore, the most efficient approach to integrate
DG and RES in a coordinated way with demand management in a sustainable system (Blanquet
et al., 2009). To take advantage from the innovative technology characteristics provided by
future smart grid, it is necessary the development of new models and control techniques which
can support and manage all the mentioned concepts and, at the same time, deal with the network
complexity and its distributed nature (Werbos, 2011).
1.3 Motivation
Currently buildings account for 40% of the world’s energy consumption and almost half of the
today’s greenhouse gas emissions. This means buildings contribute to more greenhouse gases
emissions than traffic; which is estimated at 31%. Industry is estimated at 28%. When we
breakdown and analyse building’s energy consumptions, the most worrying aspect is that most
of this energy is used for heating or cooling (StorePET, 2014).
INTRODUCTION
4
In Europe, the energy use in buildings has overall seen a rising trend over the past 20 years. For
example, in 2009, households were responsible for 68% of the total final energy use in
buildings, mainly due to space heating responsible for around 70% (Nolte, 2011). This
increasing energy consumption is mainly to fulfil the demand for thermal comfort, being
presently the HVAC systems the principal energy end use in buildings (Korolija et al., 2011).
By these facts, it is socially, environmentally and economically imperative to decrease the
energy consumption by increasing the buildings efficiency. A viable choice to achieve the
reduction of energy consumption in the building sector is the application of demand response
(DR) mechanisms. Demand Response program, is an efficient load management strategy for
customer side, it is nowadays mostly used with the encouragement of customers to shift their
habits wisely according to electricity price variation during daytime (Siano et al., 2014). The
DR potential it is not sufficiently explored, being the two key challenges to work with diverse
heterogeneous loads and with the distributed nature of renewable sources. New technology
advances in communication are providing solutions to overcome the current electricity demand
requirements. This new development has accelerated devising various industrial programs for
scheduling utilization of residential appliances (Wang et al., 2015). This DR mechanisms
combined with DSM methodologies will be relevant in the future distributed SGs (Mahmood et
al., 2014), and provide solutions that allow buildings to be fully integrated and prepared to
efficiently coexist in an dynamic and inconstant environment typically supported by renewable
resources (Figueiredo et al, 2010). Nowadays the production of electricity system follows the
load. However, the renewable sources of electricity are essentially intermittent and it is vital to
provide flexibility to the grid to absorb the variations from these sources. In the intelligent
energy system (Smart Grid) the production controls the consumption. For example, when the
wind blows or the sun shines, buildings consumption must be automatically adjusted and
consumers will go from being passive participants to be active players in the electricity system.
Several solutions are emerging to deal with the variability, flexibility and poor controllability of
the green sources and consequently on the ability to maintain the balance between demand and
supply. Remark that, DSM strategies must have into account the control of many kinds of
appliances, for instance, HVAC systems, lighting or electric vehicles charging.
The methodology here presented seeks a solution to respond to this variability, implementing,
with the technological and advanced environment that SGs are projected to provide (Paul et al.,
2014), pursuing new technologies and solutions that will allow simple home appliances to be
entirely controllable. Also, this active DSM is able manage the loads to obtain harmony in
demand supply ratio.
INTRODUCTION
5
1.4 Research Questions
With SGs, domestic customers will pass from static consumers into active participants in the
production process. Final user’s participation can be achieved due to the development of new
domestic appliances with controllable load. Shifting their electricity consumption in time or, by
changing their work conditions, these devices can be controllable, adjusting the demand to the
desired intermittent source without decreased the comfort of the residents. In distributed
networks, aggregated to an intermittent source, an unlimited number of this kind of loads can
exist, representing a control problem to achieve the network efficiencies, involving
stakeholder’s satisfaction. To incorporate this in system design and analysis, demand response
needs to be supported by scientific methodology based on analysis and models verifiable by
experiment. How to improve energy efficiency using this domestic potential is still not well
studied and needs to be a research topic.
Considering the mentioned above, this work aims providing solutions to answer the following
research questions:
Q.1 How, in a distributed network, can the demand be adjusted to an intermittent source to
maximize the energy efficiency?
Q.2 How to improve energy efficiency using the domestic potential in a distributed network?
Q.3 Which control methods should be applied in a distributed network with demand side
management to obtain all the existing energy potential from intermittent energy sources?
The adopted work hypothesis to address the research question is defined below:
Using an integrated approach, that in a distributed environment, considers multi-agent
control scheme and an optimization MPC multi-objective approach with anticipative effect,
capable to deal in a DSM perspective, with fluctuating energy sources, smart load control,
thermal comfort and real-time price negotiation.
1.5 Aimed Contributions
The work here presented is distinct and has the advantage of providing a solution that integrates
set features and concepts that are also the smart grid bases, such as intelligent control,
distributed generation, energy efficiency, energy saving, DSM, fluctuating energy sources,
smart load control, thermal comfort, real-time price negotiation and energy auction markets
mechanisms. These characteristics deliver a unique structure where the grid connected entities
actions can be intelligently combined, aiming for a more efficient, secure and sustainable energy
system. So, the work intends to provide an innovative solution involving an integrated
INTRODUCTION
6
Distributed Model Predictive Control (DMPC) to manage networks that is able to efficiently
adapt the consumption to an intermittent renewable source. It provides news developments in
DSM with renewable energy sources and backup fossil energy using an hourly auction
mechanism access and a load allocation scheme that allows managing the loads in order to
balance demand and supply. The proposed sequential multi-agent DMPC scheme for thermal
house comfort and energy savings provides robustness, adjustment and flexibility to the global
system. Also the work presents an energy usage optimization scheme with DSM in which the
consumer as the flexibility to choose by the hour between comfort or energy savings.
The proposed integrated solution provides several scientific contributions that have been
developed under the framework of this thesis:
C.1 New developments in DSM, for different thermal areas, exploit an integrated
optimization approach able to cope with intermittent renewable energy sources
and backup fossil energy using an hourly auction mechanism access - the obtained
solution allows managing the loads in order to balance demand and supply. The DSM
optimization process uses a routine that finds a constrained minimum of a quadratic cost
function that penalizes the sum of several objectives, which based on energy forecasts
and several other inputs, is able to adjust the energy consumption to an intermittent
renewable resource or even restrict it to this type of resource in detriment of fossil
sources. An auction mechanism allows stablishing an access order to the renewable
source based on an hourly energy bid. Sequentially, the controllers transmit between
them the information about the available energy;
C.2 TCA dynamic model with multi-zone dynamically coupled areas is developed -
the developed model is built for infrastructures composed by several adjacent areas that
may thermally interact among them. The dynamical model allows each room to be
treated independently, with its own construction features, thermal disturbances, comfort
requirements, energy costs and consumption needs. When distinct divisions thermally
interact the information about the predicted indoor temperature is transmitted between
them. This change of information allows to each controller a greater understanding
about the surrounding environment and consequently improving the decision-making
ability;
C.3 A novel sequential scheme for DSM based on multi-agent DMPC for thermal
house comfort and energy savings is reported - the proposed sequential scheme
provides robustness, adjustment and flexibility to the global system. The sequence is
built based on an energy bid where the highest biddings are placed first in the access
order. After consume, each agent predicts its consumption profile and pass throw the
next, the information about the predicted available renewable energy. At each hour, the
INTRODUCTION
7
sequence is established and the energy price depends from the offered bid, amount and
type (renewable or grid) of energy consumed;
C.4 Shifting load algorithm for loads allocation - customer sets features of flexible
loads and the algorithm fits them in the most favourable time interval gap. Each
customer select for each division, the characteristics of the loads, the load value, its
duration, the turned on time and the “sliding level” of the “shifted loads”. With this
information the algorithm finds all the possible load combinations and from these
selects only the feasible ones. The feasibility of the loads is related to the predicted
hourly available power, so, at each instant the look forward and tries fitting the loads in
gaps that privileges the consumption of renewable energy. Thus, in a distributed
environment with energy and comfort constraints, the algorithm is able to find the best
hour to turn on the loads;
C.5 Energy usage optimization scheme with DSM adjustable parameters - with the
developed optimization scheme the consumer has the flexibility to choose hourly
between comfort or energy savings. Each controller provides to the consumer a set of
parameterizations which according to his consumption perspective, space occupancy
schedule, indoor temperature preferences and energy cost, provides to the user the
ability to be hourly in his most convenient state.
1.6 Outline
This dissertation is structured as follows:
Chapter 1 – Introduction, here is performing a brief introduction with a background analysis,
followed by the research problem. The chapter also states the main original contributions of the
work and in the end the outline of the thesis is presented.
Chapter 2 – Literature Review and Model Based Predictive Control, the focus of this chapter
is to perform a background analysis on the main topics of the thesis, Demand Side Management,
Multi-Agent-Systems and Model Based Predictive Control. The chapter also includes general
review on distributed model predictive control formalization for complex large-scale dynamical
systems.
Chapter 3 – Dynamic Models and Scenarios Description, the main ideas and models that
support the research developed are presented. The chapter starts by presenting the general
abstract problem, the conceptual scenarios and describe the system architecture. Finally, the
developed dynamic model is addressed using a line state-space model approach.
INTRODUCTION
8
The elaboration of this chapter provided relevant matter that was originally included in part in
the following publications (Barata et al., 2014a; Barata et al., 2014b).
Chapter 4 – MPC and PI control in thermal comfort systems chapter presents the fundamental
approach towards the final developed structure. Is the first seed which allowed us to understand
the problem, its dynamics and to verify that the MPC control strategy is suitable for the
objective that is intended to be achieved. The methodologies are based on a predictive control
structure associated to a PI controller and applied to the thermal house comfort in an
environment with limited energy sources. The predictive control is compared with the
traditional control methods used in thermal systems. Also, it was created a temperature “comfort
zone” to weight only the temperatures outside the gap, instead of traditional system were all
deviations are weight. Simulation results and analysis also complement two distinct situations
where the system reacts differently if the indoor temperature is inside or outside the “comfort
zone”.
The elaboration of this chapter provided relevant matter that was originally included in part in
the following publications (Barata et al., 2012a).
Chapter 5 – Thermal Comfort with Demand Side Management Using Distributed MPC, this
chapter presents a DMPC methodology for indoor thermal comfort which simultaneously
optimizes the consumption of a limited shared energy resource. Firstly, the global scenario and
all the made assumptions that support the thesis are described. Also, the implemented sequential
architecture scheme is pictured. Secondly, the control objective is defined and the tailored
DMPC optimization problem is detailed, formalized and exemplified.
The Algorithm I which describe implemented sequential scheme is presented. This algorithm
represents the basic structure for more complex schemes.
Chapter 6 – DMPC for Thermal House Comfort with Sequential Access Auction, this chapter
presents the obtained results achieved with the developed integrative methodology to manage
energy networks from the demand side with strong presence of intermittent energy sources and
with energy storage in house-hold or car batteries.
The results are pictured with and without batteries support and two different approaches are
considered, the energy split based on a fixed sequential order, and with the energy split varying
hourly based on a bid value directly related with the consumption profile behaviour. The
Algorithm II, which describes implemented sequential scheme with variable hourly sequence
and energy storage, is also presented.
The elaboration of this chapter provided relevant matter that was originally included in part in
the following publications (Barata et al., 2012b; Barata et al., 2013a).
INTRODUCTION
9
Chapter 7 – DMPC for Thermal House Comfort with Sliding Load chapter presents the
developed DMPC integrative solution which is able to, in a distributed network with multiple
TCAs, adjust the demand to an intermittent limited energy source, using load shift and
maintaining the indoor comfort. The Algorithm III which describes the implemented shifting
and loads allocation scheme is presented followed by the results and its analysis. The
elaboration of this chapter provided relevant information that was included in the following
publications (Barata et al., 2013c) and more detailed in (Barata et al., 2014c).
Chapter 8 – Conclusions and Future Work Directions, in this last chapter, discusses the main
conclusions of this work, detailing the problems found along the way and pointing out potential
solutions for them. Additionally, the main results and contributions are summarised, and some
possible directions for further research are mentioned.
INTRODUCTION
10
11
Chapter 2
2 Literature Review and Model Based
Predictive Control
2.1 Introduction
Future SGs are expected to include distributed generation, demand side management, intelligent
control, energy efficiency, intermittent RES, thermal comfort, load control, real-time price
negotiation and energy saving (IEA, 2011). These concepts are the smart grid bases, where the
actions of all agents connected to the electricity system can be intelligently integrated aiming for
a sustainable, efficient and secure energy supply system.
Being an actual theme, and despite the present trends and R&D efforts, SGs are now giving the
firsts steps in the implementation phase, Figure 2.1, with small projects and grid prototypes all
over the world (EcoGrid, 2014; DOE, 2014). More research is needed to offer solid solutions to
overcome all the constraints and turn real this ambitious vision.
In a scenario with strong presence of intermittent RES that supply a set of houses, it is possible
to manage several loads in each house in order to: 1) adjust the demand to the supply; 2)
maintain indoor thermal comfort; 3) achieve lower energy costs 4) reduce CO2 emissions. Using
a demand side management approach, the distributed loads can be manage by negotiating in real
time the energy price and the consumer comfort in order to guarantee the balance between
demand and the intermittent source. The main goal is to develop new control methods and
methodologies to manage future SGs with demand side management with high penetration of
intermittent resources.
LITERATURE REVIEW
12
Figure 2.1. Smart Grid Roadmap (Bsria, 2014)
The method uses the match results between demand and supply to decide the control actions to
be implemented on the demand side. It is considered environmental and pre-establish variables
as a decision-making factor. By using this active DSM control, the optimal control strategies for
various appliances can be generated whilst maximum utilisation of energy supplied from
intermittent systems is guaranteed. In a distributed network, with several loads at different
locations control actions applied for solving a local energy problem can create lack of energy at
a different location in the network. Therefore, coordination control strategies are required to
make sure that all available control actions serve the same objective. To support the idea, it is
consider multi-agent control schemes in which each agent will employ a model-based predictive
control approach. The agents communicate and negotiate in a distributed network environment,
to improve decision making, and, by adjusting the demand to the source, the maximum energy
potential existent in the intermittent source can be achieved.
Figure 2.2. Venn diagram with implemented technologies.
MBPC
DSMMAS
LITERATURE REVIEW
13
At the same time, as referred, is also desirable no or negligible impact on end users/occupiers.
Thus, this literature review intend to show what are the main research trends related with the
three main support ideas that interact in this work, DSM, Multi-Agent Systems (MAS) and
Model Based Predictive Control (MBPC), Figure 2.2.
2.2 Demand Side Management for Distribution Networks
To reach important reduction in CO2 emissions, the RES will be the main contributors to the
electricity generation systems. Presently, the solar and wind power, are the renewable energy
technologies that are commercially available with sufficient scalable. Although the expected
growth of these sources through to 2020 target is envisioned, there are real concerns about the
variability, flexibility and poor controllability of the sources and consequently on the ability to
maintain the balance between demand and supply. The growing uncertainty in power systems
coupled with the introduction of power markets calls for the development of new tools for
planning, operations and market-based decision-making. To deal with this uncertainty due to the
dissemination of renewables, the systems will need, for example, to apply bigger amounts of
reserve that is generally provided by a combination of synchronised and standing reserve.
Besides to synchronise reserve provided by part-loaded plant, the balancing task will also be
supported by standing reserve, which is supplied by a plant with higher fuel and environmental
costs, such as Open Cycle Gas Turbine (OCGT), or through more desirable techniques such as
storage facilities or DSM.
Demand Side Management was introduced in USA by Electric Power Research Institute (EPRI)
in the 1980s (Balijepalli et al., 2011) and has been traditionally seen as a means of reducing
peak electricity demand so that energy suppliers could diminish the construction of new
capacity. Therefore, to decrease peak load, DSM is mainly applied via price incentives, with the
existence of lowering tariffs at times when the aggregate demand is expected to be below
average to encourage the end user to shift flexible loads towards those periods. By reducing the
overall load on an electricity network, DSM provides numerous benefits: increasing system
reliability, reducing of the dependency on energy importation, reducing energy prices,
stimulates energy markets competitively, reducing harmful emissions to the environment and in
avoiding high investments in generation, transmission and distribution networks. Thus DSM
applied to electricity systems provides significant financial, reliability and ecological benefits.
With the power markets development, new emerging technologies and the new power grid
concepts (microgrid, smart grid), innovative solutions are giving more relevance to DSM
strategy. Generally, DSM programs are separate in these four categories:
LITERATURE REVIEW
14
Energy efficiency, the user decreases the demand for energy without sacrificing the
benefits received from energy (e.g. installing building insulation, purchasing more
efficient appliances);
Conservation, the consumer decreases his energy demand by reducing their energy
usage (e.g. adequate thermostat temperatures, turning off lights);
Load management, the energy demand is reduced during periods of peak demand
when capacity is limited and the cost of energy provision is high;
Public information, which encourages customer participation in energy efficiency,
conservation, and/or load management activities through public campaigns, direct to
customer communication, or increasing customer access to information about their
consumption of energy services.
DSM directly benefits utilities in the following ways:
Distribution - only utilities avoid having to purchase additional peaking and base-load
energy resources. From the volatile wholesale energy market;
Electricity generating utilities avoid the cost of securing fuel and pollution abatement
for peaking and base-load power plants, while deferring expensive investments in new
power plants and their associated compliance costs;
Both kinds of utilities avoid costly investments in new transmission and distribution
infrastructure.
Utilities may in turn pass these savings on to consumers, resulting in lower utility bills. In
addition, DSM directly benefits utility customers in the following ways:
Many DSM programs provide financial incentives (such as rebates, bill credits, lower
rates, or low interest financing) to encourage customers to make choices that reduce
their energy consumption overall or during periods of peak demand;
By encouraging customers to reduce their energy usage or to consume energy during
times when energy services are less costly, DSM programs help customers to reduce
their monthly utility bill.
Demand response is a class of DSM programs in which electricity companies offer incentives to
clients to reduce their demand for electricity during periods of critical system conditions or
periods of high market power costs. DR programs are normally classified according to the
customer motivation method and the criteria with which load reduction “events” are triggered or
initiated. So, when the utility offers customers payments for reduction of demand during
specified periods, the program is called load response. But, when customers voluntarily reduce
their demand in response to forward market prices, the program is called price response.
Customers reduce load during those periods when the cost to reduce load is less than the cost to
LITERATURE REVIEW
15
generate or buy the energy. In this context, brief descriptions of these DR techniques that are
presently worldwide being applied and studied are now described.
Load-Response Programs
In load-response programs, utilities offer customers payments for reducing their electricity
demand for specified periods of time.
Direct load control
Direct load control is applied to domestic appliances that can be switched on and off during
periods of time using remote appliance controllers. The most common applications are the ones
that account for the most significant portion of energy demand, HVAC systems and water
heaters;
Load limiters
Load limiters have the intent to limit the power that can be used by consumers. This approach
temporarily disconnects parts of the installation, most often the HVAC or electrical heating
during power peaks. This scheme can also provide some flexibility to users to decide which
appliances to use now and what can be delayed;
Interruptible and Curtailable Load (I&C)
Interruptible and Curtailable-load programs are relatively simple to implement, the customers
agree to decrease or turn off pre-established loads for a period of time when notified by the
utility. Clients can switch off loads or adjust settings. Utilities must notify customers before an
interruptible/curtailable load event. As reward, participants receive lower electricity bills during
normal operation, as well as additional incentives for each event;
Frequency regulation
System frequency is the direct measure of the balance between generation and system demand
at any one instant and must be maintained continuously within narrow statutory limits.
Recently, there have been initiatives to investigate a technology that can be incorporated into
electrical appliances to provide frequency regulation (Dynamic Demand, 2014);
Scheduled Load
Scheduled load reductions are pre-planned between the utility and customer. Clients receive bill
reductions, as well as significant advance notification, since contractual agreements are set
months ahead. The advantage of this program is that customers can plan to reduce load on pre-
determined days;
LITERATURE REVIEW
16
Price-Response Programs
These programs allow customers to voluntarily reduce their demand in response to economic
signals.
Time of use pricing
Time of use (TOU) tariffs are designed to more closely reflect the production and investment
cost structure, where rates are higher during peak periods and lower during off-peak periods.
Basically, instead of a single flat tariff for energy use, TOU tariffs are plan to be higher when
electric demand is higher, meaning that when you use energy is as important as how much you
use. These programmes are commonly practiced in a large number of countries, particularly for
households with electric heating;
Dynamic/Real Time Pricing (RTP)
The dynamic imposed by the existing deregulated market is based on real time system of supply
and demand. Prices change time to time and hour to hour depending upon these two factors. By
exposing customers to RTP i.e. time-varying prices, they can have a better view of the
prevailing market and the information and incentive to reduce their demand at peak times or
critical peak prices (CPP) and by this way provide a financial incentive for participants to shift
optional activities, such as clothes washing and drying and dishwashing, to times other than
critical peak periods;
Demand bidding
Demand bidding programmes are available when the customer is willing to negotiate is
consumption needs in detriment of cost reductions. So, the client is disposed to decrease or
sacrifice his consumption of electricity at a certain predetermined price. For example, this
technology can be applied to a smart thermostat which controls the indoor comfort equipment’s.
So, depending from real time electricity price levels, the thermostat can be programed to allow
different sceneries in different schedules.
In buildings, novel solutions that integrate occupant behaviour within the building in relation to
energy consumption are emerging (Virote & Neves-Silva, 2012).
Only with smart metering and smart appliances technology the mentioned programs are viable.
With smart meters consumer can see exactly what power loads are driving up their energy use
and make specific changes, (Navetas, 2014). This information combined with time flexible
smart electrical appliances (HVAC systems, water heaters, refrigeration, lighting, etc.) can lead
to much better energy efficiency.
LITERATURE REVIEW
17
As showed, DSM will have an important role in the future smart grids architecture (Chen et al.,
2010; Luo et al., 2010) and by this reason many studies are being performed involving the
mentioned DSM techniques to provide future implementation solutions.
Callaway (2009), with the goal of delivering services such as regulation and load following,
developed new methods to model and control the aggregated power demand from a population
of thermostatically controlled loads (TCLs). The researcher manipulates the temperature set
point in order to control the demand curve and adapt it to a wind plant power source. The work
showed that it is possible the demand to follow the source with small changes in the nominal
thermostat temperature set point. With similar objective, Jun (2009), developed an algorithm
capable of controlling loads based on the available supply at certain time. The objective of the
algorithm is to minimize the impact on users and at the same time maximize the match between
demand and supply through identifying the best combinations of different demand side control
options such as load shifting, load control (on/off control and proportional control) and load
recover for various demand loads.
A potential application for controllable domestic heat loads, and flexible distributed generation
in power systems with significant capacities of uncertain wind generation is described in
(Savage et al, 2008). Heat load is effectively controlled by remote adjustment of thermostat set
points. The system transmit a signal to reduce heating load set points when net demand is
greater than the forecast value and a reserve shortfall was imminent. Zaidi et al., (2010) follows
a DSM approach, where unessential loads get selectively disconnected from the grid in an
under-generation scenario. In order to automatically detect unessential loads, load recognition
on the basis of measured consumption data can be performed. With the same objective of peak
shaving and balance between demand and supply, (Molderink et al., 2010), shows that with the
use of good predictions, in advance planning and real-time control of domestic appliances,
better matching of demand and supply can be achieved. With a similar approach, several other
studies can be found (James & Jones, 2010; Krishnappa, 2008).
As presented, several methods are focus in the development of load control manipulation
models, however, DSM is also been studied via models that use electricity incentive prices to
promote load management (Yang et al., 2006; Hamidi et al., 2008; Aalami et al., 2008; Yu &
Yu 2006; Shaikh & Dharme, 2009).
The desire DSM methodology that is intended to be achieved is a mix of these two solutions. It
intends to take advantage from the innovative technology characteristics provided by future
smart grids. It is expected that every electric house appliance will be controllable and the
communication infrastructures will allow RTP arrangement. At the same time, is desirable to
maximize the efficiency of renewable energy resources and minimize the consumer energy
LITERATURE REVIEW
18
costs. New models and control techniques must be developed to be capable of not only to
manage loads based on the available supply at a certain time, but also by negotiating in real time
the energy price and the consumer comfort (without significantly compromising the user
satisfaction) in order to guarantee the balance between demand and the intermittent source.
In a scenario with strong presence of intermittent RES that supply a set of houses, the new
solution must include smart devices, energy management systems (EMS) such as programmable
controllable thermostats and/or PEV charge, capable of making intelligent decisions based on
smart prices and users comfort.
2.3 Model Based Predictive Control
The firsts predictive controllers concepts were born from optimal control methodologies, such
as the Linear Quadratic (LQ) or the Linear Quadratic Gaussian (LQG) controllers in the
early’70s (Anderson and Moore, 1971). Only in late 70's that the first two truly MPC control
laws arise, the Identification-command (IDCOM) and the Dynamic Matrix Control (DMC),
which are acknowledged as the roots of MPC. These two methods share some common features
which established the basis of MPC.
The predictive control based on model MBPC is nowadays as one of the most popular and
efficient control strategies in industry. So, during the last two decades a growing interest has
been granted to model predictive control (MPC) due to its ability to handle constraints in an
optimal control environment (Moros an et al., 2010; Mosca, 1995). It is expected that soon MPC
may substitute the majority of the classic controllers that are becoming inefficient in complex
environments.
MPC is a form of control in which the current control action is obtained by solving on-line, at
each sampling instant, a finite horizon open-loop optimal control problem, using the current
state of the plant as the initial state, the optimization yields an optimal control sequence and the
first control in this sequence is applied to the plant. The major advantages that have contributed
for the success of this method are (Orukpe 2005; Jimenez 2000):
It handles multivariable control problems naturally, Single-Input/Single-Output SISO)
and Multiple-input/Multiple-Output (MIMO) formulations are similar;
Difficult dynamics such as dead-times, unstable or non-minimum phase systems can be
easily handled;
Feedforward compensation of measurable disturbances can be introduced in a natural
way exploiting the model-based and predictive features of the MPC methodology by
using disturbance models;
LITERATURE REVIEW
19
It can take account of actuator limitations The incorporation of constraints in the
manipulated and controlled variables and/or states is a simple task. Constraints can be
considered at the controller design stage and the resulting optimisation problem can
often be solved using standard Linear Programming (LP) or Quadratic Programming
(QP) tools. It also allows operation closer to constraints, hence increased profit;
MPC controllers have been developed either for linear or non-linear models;
Methods which guarantee the stability of the closed-loop system are available;
Robustness features can be enhanced through tuning parameters or optimisation
methods. Constraint handling is, indeed, one of the most appealing properties of MPC,
since limits of several kinds always occur in practice.
Constraints can be used to describe several security limits, physical restrictions, technological
requirements and so on. These requirements must be handled at the controller design stage to
avoid undesirable performance. Constraints have two main objectives. They can be used to
increase the accuracy of the model, incorporating on it the actuator and plant limits and also be
used as tuning knobs to describe control requirements or specifications. Typically, optimal
operating point lies close to (or on) one or several limits and therefore, representing an
advantage from an economical point of view when it operate as close to the constraint boundary
as possible. Constraints can be classified according to different criteria. The following
classification of constraints, according to practical considerations, is due to Álvarez and de
Prada (1997):
Physical constraints. These limits, which must never be surpassed, are determined by
the physical limitations of the system;
Operating constraints. These bounds are fixed by the plant operators to specify the
optimal operating region. The operation constraints are more restrictive than he physical
limits;
Optimization or set point conditioning constraints. These limits, more restrictive than
the operating constraints, are used only if the set point conditioning technique is
applied;
Working constraints. These are the actual constraints considered by the controller to
determine the feasible region. The working constraints are obtained by choosing the
most restrictive among the physical, the operating and the optimization limits. Apart
from constraint handling, the relevant issues of stability and robustness have been
successfully tackled in the last decade.
The MPC block diagram is present in Figure 2.3 where the main blocks represent:
LITERATURE REVIEW
20
Reference Trajectory - represents the desire signal to future outputs. The previous knowledge
of this trajectory provides the anticipative effect of this controller;
Model - the mathematical model, linear or nonlinear, that represents the process must be
capable to describe its dynamic behavior with sufficient accuracy;
Predictor- trough the mathematical model provides the future outputs based on the actual plant
information;
Optimizer - minimizes the cost function at every time sample to obtain the control action that
guaranties the system performance. When linear models are used and the problem is
unconstraint, the quadratic cost function presents a analytical solution, otherwise, some numeric
optimization methods must be used.
As can be seen, the presence of the plant model is a necessary condition for the development of
the predictive control. The success of MPC depends on the degree of accuracy of the plant
model (Zheng 2010).
Figure 2.3. Basic block diagram of MPC
Predictive control uses the receding horizon principle. This means that after computation of the
optimal control sequence, only the first control sample will be implemented, subsequently the
horizon is shifted one sample and the optimization is restarted with new information of the
measurements. The prediction horizon remains the same length despite the repetition of the
optimization at future instants. Since the state prediction x and hence the optimal sequence u*
depend on the current state measurements 𝑥(𝑘), this procedure introduces feedback into the
MPC law, providing a degree of robustness to modelling errors and uncertainty. Figure 2.4
explains the idea of receding horizon.
OPTMIZER
MODEL
(ESTIMATOR)
PROCESS
PREDICTOR
control
Predicted
error
Predicted
output
output +
–
REFERENCE
TRAJECTORY
COST
FUNCTION CONSTRAINTS
u(k)
𝑦 (𝑘 + 1)
r(k)
y(k)
𝑥 (𝑘)
LITERATURE REVIEW
21
Figure 2.4. Conceptual picture of the moving horizon in predictive control (Boom &
Stoorvogel, 2010).
In Figure 2.4 Nm, NC and HP represents is the minimum cost horizon, the control horizon and
prediction horizon respectively. At time k the future control sequence {𝑢(𝑘|𝑘), . . . , 𝑢(𝑘 + 𝑁𝑐 −
1|𝑘)} is optimized such that the performance-index J(u, k) is minimized subject to constraints.
At time k the first element of the optimal sequence (𝑢(𝑘) = 𝑢(𝑘|𝑘)) is applied to the real
process. At the next time instant the horizon is shifted and a new optimization at time k+1 is
solved.
The goal of MPC is to minimize a cost function over a given prediction horizon period. This
cost function should give an indication for the performance of the system. The control signal
contains the setting for the control measures that are able to influence the system, and the
constraints may contain the upper and lower bounds on the control signal, but also linear or non-
linear equality and inequality constraints on the control inputs and the states of the system.
In MPC, control decisions, system inputs, u(k) are made at discrete time instants k = 0,1,2,...,
which usually represent equally spaced time intervals. At decision instant k, the controller
samples the state of the system x(k) and then solves an linear optimization problem of the
following form to find the control action:
𝑚𝑖𝑛𝑢 𝐽(𝑥, 𝑢; 𝑥(𝑘)),
(2.1)
𝑠. 𝑡. 𝑥(𝑘 + 1) = 𝐴𝑥(𝑘) + 𝐵𝑢(𝑘), (2.2)
𝑢 ∈ 𝒰,
where x(k) ℝn and u(k) ℝm, and A ℝn×n, B ℝn×m, are known matrices with constant
entries of corresponding entries. The cost function J(k) is linear quadratic, Linear Quadratic
Predictive Control (LQPC), with symmetric weighting matrices Q and R, where Q ≥ 0, R > 0,
and 𝒰 is an admissible control set.
HP
LITERATURE REVIEW
22
𝑚𝑖𝑛𝑢0,…,𝑢𝐻𝑃−1
= ∑ 𝑥(𝑘)𝑇𝑄𝑥(𝑘) + 𝑢(𝑘)𝑇𝑅𝑢(𝑘)
𝐻𝑃−1
𝑘=0
, (2.3)
𝑠. 𝑡. 𝑥(𝑘 + 1) = 𝐴𝑥(𝑘) + 𝐵𝑢(𝑘), 𝑘 = 0,…𝐻𝑃 − 1, (2.4)
𝑢(𝑘) ∈ 𝒰, 𝑥(𝑘 + 1) ∈ 𝒳, 𝑘 = 0,… , 𝐻𝑃 − 1,
𝑥0 = 𝑥(𝑡) ≡ 𝑥(𝑘).
The future states are described as,
��(k) = 𝐺𝛥𝑈(𝑘) + 𝛤𝑥(𝑘), (2.5)
��(𝑘) = [𝑥 (𝑘 + 1)⋮𝑥 (𝑘 + 𝐻𝑃)
] = [𝐴⋮𝐴𝐻𝑃
]⏟ 𝛤
𝑥(𝑘) + [𝐵 ⋯ 0⋮ ⋱ ⋮𝐴𝐻𝑃−1𝐵 ⋯ 𝐵
]⏟
𝐺
[𝛥𝑢(𝑘)⋮𝛥𝑢(𝑘 + 𝐻𝑃 − 1)
]
⏟ 𝛥𝑈
, (2.6)
��0 = 𝛤𝑥 (𝑘). (2.7)
For regulation proposes, and considering the general case which intent to lead the state to zero,
the linear MPC without constraints solution of (2.3) is given by:
𝛥𝑈∗ = −𝐾𝑐𝑥 (𝑘) = −(𝐺𝑇𝑄𝐺 + 𝑅)−1𝐺𝑇𝑄��0 (2.8)
and implementing the first element of the optimal prediction at each sampling instant k defines a
receding horizon control law
𝑢(𝑘) = 𝛥𝑈∗(1). (2.9)
Remark that if 𝒰 and 𝒳 are polyhedral (a finite intersection of half-space), this is a Quadratic
Program, that can be can be solve fast, efficiently and reliably, using modern solver tools.
2.3.1 Linear Plant Model
For linear systems, the dependence of predictions x(k) on u(k) is linear. A quadratic
minimization cost function as (2.3) is therefore a quadratic function of the input sequence vector
u. Thus J(k) can be expressed as a function of u in the form,
𝐽(𝑘) = 𝒖𝑇(𝑘)𝐻𝒖(𝑘) + 2𝑓𝑇𝒖(𝑘) + 𝑔, (2.10)
where H is a constant positive definite (or possibly positive semi-definite) matrix, and f, g are
respectively a vector and scalar which depend on x(k). The online MPC optimization therefore
comprises the minimization over u of a quadratic objective:
LITERATURE REVIEW
23
𝑚𝑖𝑛𝑢 𝒖𝑇(𝑘)𝐻𝒖(𝑘) + 2𝑓𝑇𝒖(𝑘). (2.11)
In the absence of constraints, unconstrained MPC, the optimization u*(k) = 𝑚𝑖𝑛𝐮 𝐽(𝑘) has a
closed-form solution which can be derived by considering the gradient of J with respect to u:
𝛻𝑢 𝐽 = 2𝐻𝒖 + 2𝐹𝑥(𝑘). (2.12)
Clearly 𝛻u𝐽 =0 must be satisfied at a minimum point of J, and since H is positive definite (or
positive semidefinite), any u such that 𝛻u𝐽 =0 is necessarily a minimum point. Therefore the
optimal u* is unique only if H is non-singular and is then given by,
𝒖∗(𝑘) = −𝐻−1𝐹𝑥(𝑘). (2.13)
Remark that, sufficient conditions for H to be non-singular are for example that R > 0 or Q, �� >
0 and the pair (A,B) is controllable.
If H is singular (i.e. positive semi-definite rather than positive definite), then the optimal u* is
non-unique, and a particular solution of 𝛻u𝐽 =0 has to be defined as,
𝑢(𝑘) = −𝐻†𝐹𝑥(𝑘), (2.14)
where H† is a left inverse of H (so that H†H = I).
Implementing the first element of the optimal prediction u*(k) at each sampling instant k defines
a receding horizon control law. Since H and F are constant, this is a linear time-invariant
feedback controller u(k) = KC x(k), where the gain matrix KC is the first row of −H−1F for the
single-input case (or the first nu rows for the case that u has dimension nu), i.e.
𝑢(𝑘) = 𝑢∗(𝑘|𝑘) = 𝐾𝐶𝑥(𝑘), (2.15)
𝐾𝐶 = −[𝐼𝑛𝑢 0 … 0]𝐻−1𝐹. (2.16)
Constrained MPC is used when structural limitations of the system to be controlled usually
leads to control restrictions and avoiding such limitations may cause harmful effects as unstable
cycle. However, for performance purpose, the control system should try to use the system as
close to those limits as possible, without crossing them. Therefore, MPC uses a more direct
approach by finding the optimal control in such a way that constraints are not violated.
Most typically, linear input and state constraints likewise imply linear constraints on u(k) which
can be expressed
𝐴𝑐𝒖(𝑘) ≤ 𝑏𝑐 , (2.17)
LITERATURE REVIEW
24
where Ac is a constant matrix and, depending on the form of the constraints, the vector bc may
be a function of x(k). So, considering the optimization function (2.11) but now subject to linear
constraints:
𝑚𝑖𝑛𝑢 𝒖𝑇(𝑘)𝐻𝒖(𝑘) + 2𝑓𝑇𝒖(𝑘),
(2.18)
𝑠. 𝑡. 𝐴𝑐𝒖(𝑘) ≤ 𝑏𝑐 . (2.19)
This class of optimization problem is known as a quadratic programming (QP) problem, and
given that H is a positive definite matrix and the constraints are linear, it is easy to show that
(2.18) and (2.19) are a convex problem (i.e. both the objective (2.18) and constraints (2.19) are
convex functions of the optimization variable u. This kind of problem can be solved efficiently
and reliably using commercial solver and specialized algorithms.
Generally, the constraints which are applied on control input, state or output assume the
following form:
𝑢 ≤ 𝑢(𝑘) ≤ 𝑢, (2.20)
𝑥 ≤ 𝑥(𝑘) ≤ 𝑥, (2.21)
𝑦 ≤ 𝑦(𝑘) ≤ 𝑦, (2.22)
or even rate constraints:
∆𝑢 ≤ 𝑢(𝑘) − 𝑢(𝑘 − 1) ≤ ∆𝑢. (2.23)
Input constraints commonly arise as a result of actuator limits, e.g. torque saturation in d.c.
motors and flow saturation in valves and the state constraints may be active during transients,
e.g. aircraft stall speed, or in steady-state operation as a result of e.g. economic constraints on
process operation. In addition to the obvious equality constraints that the state and input should
satisfy the model dynamics, inequality constraints on input and state variables are encountered
in every control problem. While the equality constraints are usually handled implicitly (i.e. the
plant model is used to write predicted state trajectories as functions of initial conditions and
input trajectories), the inequality constraints are imposed as explicit constraints within the
online optimization problem.
However, constraints are classified as hard or soft. The hard constraints, must always be
satisfied, and when is not possible, the problem is becomes infeasible.
LITERATURE REVIEW
25
Figure 2.5. Hard constraint representation (Adapted from Froisy, 1994).
On the other hand, soft constraints may be violated if necessary to avoid infeasibility, generally
with a penalization in the cost function.
Figure 2.6. Soft constraint representation (Adapted from Froisy, 1994).
Chapter 4 shows an example of hard constraints application and in Chapter 6 and 7 soft
constraints. Thus, constraints will be here generically described and exemplified.
To implement a soft constraint, allowing for example that the output variable exceeds the
constraint value in a specified value 𝜖 ≥ 0,
𝑦(𝑡 + 𝑗) ≤ 𝑦 + 𝜖(𝑡 + 𝑗),
𝜖(𝑡 + 𝑗) ≥ 0, 1 ≤ 𝑗 ≤ 𝑁𝑠𝑜𝑓𝑡 .
with 𝑁𝑠𝑜𝑓𝑡 representing the horizon where the soft constraint is considered.
Thus a new term is included in the cost function (2.18),
𝑚𝑖𝑛𝑢 𝟏
𝟐𝒗𝑇𝑯𝒗 + 𝒃
𝑇𝒗, (2.24)
with
time
time
Maximum
constraint value
Maximum
constraint value
LITERATURE REVIEW
26
𝑯𝒖 = −𝑏, (2.25)
𝒗𝑇 = [𝒖𝑇 𝜖𝑇], 𝑯 = [𝑯 00 𝜆𝑟𝑰
],
𝒃𝑇= [𝒃𝑇 0], 𝜖 = [𝜖(𝑡 + 1) ⋯ 𝜖(𝑡 + 𝑁𝑠𝑜𝑓𝑡)].
Where λ𝑟 > 0 is the ponderation factor and 𝐇 the Hessian matrix. The optimization problem
must also include,
𝑦(𝑡 + 𝑗) − 𝜖(𝑡 + 𝑗) ≤ 𝑦, 𝜖 ≥ 0. (2.26)
Now, to be able to be solved as a quadratic optimization problem the constraints must be
expressed as inequalities constraints, assuming the following form:
𝑚𝑖𝑛𝑢 𝟏
𝟐𝒗𝑇𝑯𝒗 + 𝒃
𝑇𝒗, (2.27)
𝑠. 𝑡. 𝐴𝒖 ≤ 𝑐. (2.28)
In a systematic way other soft constraints can be considered in the output value,
𝑦(𝑡 + 𝑗) − 𝜖(𝑡 + 𝑗) ≤ 𝑦(𝑡 + 𝑗) ≤ 𝑦(𝑡 + 𝑗) + 𝜖(𝑡 + 𝑗), 𝑁𝑚 ≤ 𝑗 ≤ 𝐻𝑃 , (2.29)
And expressing the constraints as manipulated variables
𝑮𝒖 − 𝜖 ≤ 𝟏𝑦, (2.30)
−𝑮𝒖 − 𝜖 ≤ −𝟏𝑦, 𝜖 ≥ 0.
With 1 representing a (HP-Nm) matrix with unitary values, and G the dynamic matrix and the
incremental control vector.
To implement hard constraints on U, I
𝑈 ≤ 𝑢(𝑡 + 𝑗) ≤ 𝑈, 𝑁𝑚 ≤ 𝑗 ≤ 𝐻𝑃. (2.31)
Expressing the constraints as manipulated variables,
𝑻𝒖 ≤ 𝟏𝑈 − 1𝑢(𝑡 − 1), (2.32)
−𝑻𝒖 ≤ −𝟏𝑈 + 1𝑢(𝑡 − 1).
Where T is an upper triangular matrix with (NC×NC). Considering also hard constraints in y, the
output is written as,
LITERATURE REVIEW
27
𝑦(𝑡 + 𝑗) ≤ 𝑦(𝑡 + 𝑗) ≤ 𝑦(𝑡 + 𝑗), 𝑁𝑚 ≤ 𝑗 ≤ 𝐻𝑃 , (2.33)
and once again,
𝑮𝒖 ≤ 𝟏𝑦, (2.34)
−𝑮𝒖 ≤ −𝟏𝑦.
Supposing a set of constraints acting over a process, where the control signal U amplitude, the
actuator slew rate u and output signal y are limited. These constraints can be expressed as,
𝑈 ≤ 𝑢(𝑡) ≤ 𝑈 ∀ 𝑡,
𝑢 ≤ 𝑢(𝑡) − 𝑢(𝑡 − 1) ≤ 𝑢 ∀ 𝑡, (2.35)
𝑦 ≤ 𝑦(𝑡) ≤ 𝑦 ∀ 𝑡,
Considering that the process has m inputs and n outputs with constraints acting along a horizon
HP, (2.35) can be reformulated
𝟏𝑈 ≤ 𝑻𝒖 + 𝑢(𝑡 − 1) ≤ 𝟏𝑈 ∀ 𝑡,
𝟏𝑢 ≤ 𝒖 ≤ 𝟏𝑢 ∀ 𝑡, (2.36)
𝟏𝑦 ≤ 𝑮𝒖 ≤ 𝟏𝑦 ∀ 𝑡,
where 1 is a matrix (HP×n)×m with HPm×m identity matrices, and T is a lower triangular matrix
where the non-null inputs are identity matrices with (m×m), and writing the constraints in a
condensed form, results in,
𝐴𝒖 ≤ 𝑐,
with,
𝑨 =
[
𝑻−𝑻
𝑰𝐻𝑃×𝐻𝑃−𝑰𝐻𝑃×𝐻𝑃
𝑮−𝑮 ]
; 𝒄 =
[ 𝟏𝑈 − 𝟏𝑢(𝑡 − 1)−𝟏𝑈 + 𝟏𝑢(𝑡 − 1)
𝟏𝑢−𝟏𝑢
𝟏𝑈−𝟏𝑈 ]
. (2.37)
The QP solution can be obtained using a modified Lemke’s method (Camacho, 1993; Igreja and
Cruces, 2002).
LITERATURE REVIEW
28
2.3.2 Nonlinear Plant Model
Although the majority of processes are inherently nonlinear, most of the applications use linear
dynamic models. Several reasons are behind this fact, or because empirical linear models can be
easily identified because mostly process work nearby a desire operation point, or because linear
models may present sufficient precision when in presence of high quality the feedback
measures. Linear models may use quadratic objective functions that are simply solve with LQ
convex programming which provide a solution that should converge to the optimal criterion.
The use of nonlinear models in MPC is motivated by the need to improve the control of
processes with robust linearity’s through by improving the prediction quality.
Thus, MPC has a large population of potential applications even for nonlinear systems,
according to (Allgöwer et al., 2004), the key characteristics and properties of NMPC are:
NMPC allows the direct use of nonlinear models for prediction;
NMPC allows the explicit consideration of state and input constraints;
In NMPC a specified time domain performance criteria is minimized on-line;
In NMPC the predicted behaviour is in general different from the closed-loop
behaviour;
For the application of NMPC typically a real-time solution of an open-loop optimal
control problem is necessary;
To perform the prediction the system states must be measured or estimated;
The optimization problem is nonconvex, more difficult to solve than QP;
The computation time increases because the difficulty to find the solution of the
optimisation problem;
If a nonlinear prediction model is employed, then due to the nonlinear dependence of the state
predictions x(k) on u(k), the MPC optimization problem is significantly harder than for the
linear model case. This is because the cost in equation (2.3), which can be written as J(u(k),
𝑥(𝑘)), and the constraints, g(u(k), x(k)) ≤ 0, are in general nonconvex functions of u(k), so that
the (2.3) optimization problem,
𝑚𝑖𝑛𝒖 𝐽(𝒖; 𝑥(𝑘)),
(2.38)
𝑠. 𝑡. 𝑔(𝒖; 𝑥(𝑘)), (2.39)
becomes a nonconvex nonlinear programming (NLP) problem. As a result there will in general
be no guarantee that a solver will converge to a global minimum of (2.38) and the times
required to find even a local solution are typically orders of magnitude greater than for QP
LITERATURE REVIEW
29
problems of similar size. Unlike QP solvers, the computational loads of solvers for nonlinear
programming problems are strongly problem-dependent.
2.3.3 Generalized Predictive Control
Another similar MPC strategy that will be used in Chapter 4, is Generalized Predictive Control.
(GPC). GPC was developed by (Clarke et al., 1987) and is a popular form of MPC. In the
original version of GPC, stability was not guaranteed because of using finite horizon. New
versions of GPC were based on the idea that the stability of GPC could be guaranteed if in the
last part of the prediction horizon the future outputs are constrained at the desired set point, and
also the prediction horizon is properly selected. The well-known example for this category is
Stabilizing Input/Output Receding Horizon Control (SIORHC). The GPC controller uses a
Controlled Auto-Regressive Integrated Moving Average (CARIMA) model as presented in the
next difference equation,
𝐴(𝑞−1)𝛥𝑦(𝑡) = 𝑞−𝑑𝐵(𝑞−1)𝛥𝑢(𝑡) + 𝐶(𝑞−1)𝜉(𝑡), (2.40)
where y(t) is the process output, u(t) is the control variable, d is the delay, (t) is the measure
noise (perturbations, model errors). The GPC control law is obtained with the minimization of
the following equation,
𝐽𝐺𝑃𝐶 = ∑ [𝑦(𝑡 + 𝑗) − 𝑦𝑟(𝑡 + 𝑗)]2
𝐻𝑃
𝑗=𝑁𝑚
+∑𝑅𝛥𝑢2(𝑡 + 𝑗 − 1),
𝑁𝐶
𝑗=1
(2.41)
where R is the control weight, Nm is the minimum horizon, specifying the beginning of the
horizon in the cost function from which point the output error is taken into account, and HP
represents the prediction horizon, specifying the last output error that is taken into account. NC
is the control horizon, y(t+j) and yr(t+j) are the output and reference signal, u(t+j–1) is the
control signal increment at (t+j–1). The control weight and control horizon are the main two
tunings GPC parameters that allows obtaining different predictive controllers in order to adjust
them to the desire control plant (Clarke et al., 1987). Remark that, the control action affects the
process output only after the delay, it is reasonable to choose the minimum horizon higher or
equal to the process delay. If the process delay is unknown then, the delay can be set to one and
the minimum horizon to zero without the loss of stability. The choice of the minimum horizon
can be interesting in case of non-minimum phase processes.
Consider the polynomial,
𝐶(𝑞−1) = 𝐴(𝑞−1)𝛥𝐸𝑗(𝑞−1) + 𝑞−𝑗𝐹𝑗(𝑞
−1), (2.42)
LITERATURE REVIEW
30
where,
𝐸𝑗(𝑞−1) = 1 + 𝑒1𝑞
−1+ . . . +𝑒𝑛𝑒𝑞−𝑛𝑒 ,
𝐹𝑗(𝑞−1) = 𝑓0 + 𝑓1𝑞
−1+ . . . +𝑓𝑛𝑓𝑞−𝑛𝑓 ,
(2.43)
ne = j –1; nf = max(na, nc – j),
are stablished knowing j and A(q–1) e C(q–1).
By manipulating the model equation (2.40) and (2.42), equation (2.46) can be obtained,
𝑦(𝑡 + 𝑗) =𝐹𝑗(𝑞
−1)
𝐶(𝑞−1)𝑦(𝑡) +
𝐺𝑗′(𝑞−1)
𝐶(𝑞−1)𝛥𝑢(𝑡 + 𝑗 − 𝑑) + 𝐸𝑗(𝑞
−1)𝜉(𝑡 + 𝑗), (2.44)
with,
𝐺𝑗′(𝑞−1) = 𝐸𝑗(𝑞
−1)𝐵(𝑞−1). (2.45)
The noise is uncorrelated from the measurable signals in instant t, so the output prediction can
be written for (t + j) as,
𝑦 (𝑡 + 𝑗) =𝐹𝑗(𝑞
−1)
𝐶(𝑞−1)𝑦(𝑡) +
𝐺𝑗′(𝑞−1)
𝐶(𝑞−1)𝛥𝑢(𝑡 + 𝑗 − 𝑑), (2.46)
with,
𝐺𝑗′(𝑞−1) = 𝐶(𝑞−1)𝐺𝑗(𝑞
−1) + 𝑞−𝑗��𝑗(𝑞−1), (2.47)
and substituting in (2.44),
𝑦 (𝑡 + 𝑗) =𝐹𝑗(𝑞
−1)
𝐶(𝑞−1)𝑦(𝑡) +
��𝑗(𝑞−1)
𝐶(𝑞−1)𝛥𝑢(𝑡 − 𝑑)
⏟ 𝑎𝑣𝑎𝑖𝑙𝑎𝑏𝑙𝑒 𝑖𝑛𝑓𝑜𝑟𝑚𝑎𝑡𝑖𝑜𝑛 𝑎𝑡 𝑡 𝑖𝑛𝑠𝑡𝑎𝑛𝑡
+ 𝐺𝑗(𝑞−1)𝛥𝑢(𝑡 + 𝑗 − 𝑑).⏟
𝑓𝑢𝑡𝑢𝑟𝑒
(2.48)
From (2.48) the effect of the past and future control actions, can be separated,
𝑦 (𝑡 + 𝑗/𝑡) =𝐹𝑗(𝑞
−1)
𝐶(𝑞−1)𝑦(𝑡) +
��𝑗(𝑞−1)
𝐶(𝑞−1)𝛥𝑢(𝑡 − 𝑑) (2.49)
Equation (2.46) can be represented in vectorial form,
�� = 𝐺𝛥𝑈 + 𝑓 (2.50)
With f formed with the free responses,
𝑓 = [𝑦 (𝑡 +𝑁𝑚𝑡) 𝑦 (𝑡 + 𝑁𝑚 +
1
𝑡… 𝑦 (𝑡 +
𝐻𝑃𝑡)]𝑇
, (2.51)
and
LITERATURE REVIEW
31
𝛥𝑼 = [𝛥𝑢(𝑡) 𝛥𝑢(𝑡 + 1) … 𝛥𝑢(𝑡 + 𝑁𝐶 − 1)]𝑇 , (2.52)
where,
�� = [𝑦 (𝑡 + 𝑁𝑚) 𝑦 (𝑡 + 𝑁𝑚 + 1) … 𝑦 (𝑡 + 𝐻𝑃)]𝑇 , (2.53)
G =
[ 𝑔𝑁𝑚−𝑑 … 𝑔0 0 0 ⋯ 0
𝑔𝑁𝑚−𝑑+1 … 𝑔1 𝑔0 0 ⋯ 0
⋮ ⋱ ⋱ ⋮⋮ ⋱ 𝑔0⋮ … ⋮𝑔𝐻𝑃−𝑑 𝑔𝐻𝑃−𝑑−1 … 𝑔𝐻𝑃−𝑁𝐶−𝑑+1]
, (2.54)
with G (𝐻𝑃 – Nm + 1)× NC dimension.
GPC cost function can also be presented in vectorial form,
𝐽𝐺𝑃𝐶 = [�� − 𝑌𝑟]𝑇[�� − 𝑌𝑟] + 𝑅𝛥𝑈
𝑇𝛥𝑈, (2.55)
where
𝑌𝑟 = [𝑦𝑟(𝑡 + 𝑁𝑚) 𝑦𝑟(𝑡 + 𝑁𝑚 + 1) … 𝑦𝑟(𝑡 + 𝐻𝑃)]𝑇. (2.56)
Thus, (2.55) is minimized obtaining the follow control law,
𝛥𝑈 = [𝐺𝑇𝐺 + 𝑅𝐼]−1𝐺𝑇[𝑌𝑟 − 𝑓]. (2.57)
Being GPC a receding horizon controller, only the first element of the calculated control signal
sequence is to be applied on the process. The procedure of minimization of the cost function is
repeated in the next sampling instant. The applied control signal is:
𝑢(𝑡) = 𝑢(𝑡 − 1) + ∆𝑢(𝑡). (2.58)
So, GPC allows threating process with unknown or varying delays, constrained systems,
nonlinearities, non-minimum phase processes as well open-loop unstable plants (Clarke et al.,
1987b). The GPC control structure is presented in Figure 2.7.
Figure 2.7. GPC control structure.
LITERATURE REVIEW
32
With M
𝑀 = [1 0 … 0][(𝐺𝑇𝐺 + 𝛬𝐼)−1𝐺𝑇], (2.59)
where 𝑀 = [𝑚𝑁𝑚 𝑚𝑁𝑚+1 … 𝑚𝐻𝑃].
So, the control applied to the process can be written as,
𝛥𝑢(𝑡) = 𝑀[𝑌𝑟 − 𝑓]. (2.60)
According Figure 2.7, the control law is given by,
𝑅(𝑞−1)𝛥𝑢(𝑡) = 𝑇(𝑞−1)𝑦𝑟(𝑡 + 𝐻𝑃) − 𝑆(𝑞−1)𝑦(𝑡), (2.61)
where
𝑅(𝑞−1) = 𝐶(𝑞−1) + 𝑞−𝑑 ∑ 𝑚𝑗��𝑗
𝐻𝑃
𝑗=𝑁𝑚
(𝑞−1), (2.62)
𝑆(𝑞−1) = ∑ 𝑚𝑗
𝐻𝑃
𝑗=𝑁𝑚
𝐹𝑗(𝑞−1), (2.63)
𝑇(𝑞−1) = 𝐶(𝑞−1) ∑ 𝑚𝐻𝑃+𝑁𝑚−𝑗
𝐻𝑃
𝑗=𝑁𝑚
𝑞−(𝑗−𝑁𝑚), (2.64)
nr = 𝐻𝑃 – Nm; ns = max(na, nc – 𝐻𝑃) and nt = max(nc, nb).
Substituting (2.61) in (2.40),
𝑦(𝑡) =[𝑞−𝑑𝐵(𝑞−1)𝑇(𝑞−1)]𝑦𝑟(𝑡 + 𝑁2)
[𝐴(𝑞−1)𝛥𝑅(𝑞−1) + 𝑞−𝑑𝐵(𝑞−1)𝑆(𝑞−1)]+
𝐶(𝑞−1)𝑅(𝑞−1)𝜉(𝑡)
[𝐴(𝑞−1)𝛥𝑅(𝑞−1) + 𝑞−𝑑𝐵(𝑞−1)𝑆(𝑞−1)] (2.65)
With (2.65), the closed loop characteristic polynomial is calculated by,
𝑃𝑚𝑓(𝑞−1) = 𝐴(𝑞−1)𝛥𝑅(𝑞−1) + 𝑞−𝑑𝐵(𝑞−1)𝑆(𝑞−1). (2.66)
As mentioned, the first generation in the MPC family, such as DMC, IDCOM and GPC used a
finite prediction horizon. This feature made it possible to incorporate constraints in the control
strategy, a capability that is not supported by infinite horizon LQ control. Initially in GPC,
because of using finite horizon, stability was not guaranteed in his original version. But later,
this issue was overcome with the idea that the stability of GPC could be guaranteed if in the last
part of the prediction horizon the future outputs are constrained at the desired set point and the
prediction horizon is properly selected. So, the stability weaknesses of the first generation were
overcome with the second generation with these two methods that are the most popular in this
category, the SIORHC (Mosca et. al 1990), or Constrained Receding Horizon Predictive
LITERATURE REVIEW
33
Control (CRHPC) and Stable Generalised Predictive Control (SGPC). SIORCH optimize a
quadratic function over a prediction horizon subject to the condition that the output matches the
reference value over a further constraint range.
Another kind of adaptive predictive controller algorithm is Multistep Multivariable Adaptive
Regulator, MUSMAR (Greco et al., 1984) is a data driven algorithm designed for the adaptive
regulation of linear Autoregressive–Moving-Average Model with exogenous inputs model
(ARMAX) plants. This algorithm combines several features, the optimization of a receding
horizon quadratic cost, it assumes a constant feedback over the prediction horizon, the cost
function is minimized based on predictive models for both plant input and output signals and the
estimation of the predictors and the data are separate.
This algorithm possesses attractive local self-optimizing properties and was extendedly studied
during the eighties and nineties years of the last century (Mosca at al., 1989; Rato et al., 1997)
and successfully applied in recent years in many areas (Nunes, et al., 2007).
2.3.4 Stability and Feasibility
In this work, feedback stability is provided by choosing a sufficiently long predictive horizon
and proven by results. Feasibility is achieved by the use of soft constraints (5.1) in the
optimization problem formulation.
Table 2.1. Simulation parameters for the several prediction horizons
Parameter A1 Units
𝛯 1
5000
Φ 200
1
Req 50 ºC/kW
Ceq 9.2103 kJ/ºC
To exemplify, several predictive horizons were chosen to demonstrate the influence of this
parameter. The simulation time was also counted to evaluate the system performance under the
several HP chosen, 1, 6, 12 and 24h. The simulations were made for one house in the same
conditions used in Chapter 5, using optimization problem (5.1) and the dynamical model (3.1).
The used parameters are described in Table 2.1 and the obtained results in Figure 2.9 to Figure
2.11. It is considered that the system knows the existence of load disturbances, Figure 2.8,
inside the selected HP. Simulations were made in an Intel Core 1.59GHz, 2,0GB of RAM
machine.
LITERATURE REVIEW
34
Figure 2.8. Disturbance load profile in the several prediction horizons.
Figure 2.9. Indoor temperature profile for the several prediction horizons.
Figure 2.10. Power profile for the several prediction horizons.
0 5 10 15 20-0.5
0
0.5
1
1.5
2
Time (h)
Pow
er
(kW
)
Disturbance Load (kW)
0 5 10 15 2021
22
23
24
25
26
27
28
Time (h)
Tem
pera
ture
(ºC
)
Temp. Max.
Temp. Min.
Indoor Temp HP=1.
Indoor Temp HP=6.
Indoor Temp HP=12.
Indoor Temp HP=24.
0 5 10 15 20-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
Time (h)
Pow
er
(kW
)
Power Max.
Power Min.
Control Input HP=1
Control Input HP=6
Control Input HP=12
Control Input HP=24
LITERATURE REVIEW
35
Figure 2.11. Consumption profile for several prediction horizons.
As can be seen in Figure 2.9, with HP=1 the system becomes instable, the indoor temperature
diverge from the chosen comfort zone because with this horizon the power consumption is
almost null. The indoor temperature and power profiles are similar with the others HP, but
remark that, with HP=24, the system is able to anticipate sooner all the disturbances and
constraints and by that reason the power consumption is higher in order to also maintain the
temperature inside bounds. It can be seen that the compromise between comfort and
consumption is notorious, the controllers are always attempting to accomplish both constraints,
but when is not possible, it compensates with one inside range and the other outside. The
application of soft constraints allows this behaviour and the problem feasibility.
As expected, Table 2.2 shows that higher prediction horizons has the drawback of higher
computational efforts, the compromise between anticipation, feasibility and computational
efforts must be taken in account when the parameters are chosen.
Table 2.2. Computational efforts under several prediction horizons
HP (h) Computational effort (s)
1 2,8
6 4,4
12 52,5
24 289,5
2.3.5 Centralized MPC
The continuous instrumentation development associated to the need of improvements in
processes and environmentally conscious solutions has made the process control become even
0 5 10 15 200
2
4
6
8
10
12
14
Time (h)
Pow
er
(kW
h)
HP=1
HP=6
HP=12
HP=24
LITERATURE REVIEW
36
more complex. With this, the implementation and maintenance of centralized MPC algorithms
have become an important issue and often complicated to treat. Some of the major obstacles
encountered in the use of advanced control techniques, such as MPC, are the difficulty of
communication between sensors and processing units in geographically distributed systems and
computational limitations for some optimization problems. Although, for large and
geographically distributed systems, the two mostly used control strategies are centralized or
decentralized. Centralized control can coordinate subsystems but may suffer from the curse of
dimensionality and failures in the central control unit. In general, in most of the applications,
MPC is implemented in a centralized scheme, where the complete system is modelled and all
the control inputs are computed in one optimization problem, Figure 2.12.
Figure 2.12. Centralized MPC architecture (Scattolini, 2009).
The overall system model can be represented as a discrete, linear time invariant (LTI) model
with the form,
𝑥(𝑘 + 1) = 𝐴𝑥(𝑘) + 𝐵𝑢(𝑘), (2.67)
𝑦(𝑘) = 𝐶𝑥(𝑘) ,
in which,
𝐴 =
[ 𝐴11 𝐴12 … 𝐴1𝑁𝑠⋮ ⋮ ⋱ ⋮𝐴𝑖1 𝐴𝑖2 … 𝐴𝑖𝑁𝑠⋮ ⋮ ⋱ ⋮
𝐴𝑁𝑠1 𝐴𝑛𝑠2 … 𝐴𝑁𝑠𝑁𝑠]
, 𝐵 =
[ 𝐵11 𝐵12 … 𝐵1𝑁𝑠⋮ ⋮ ⋱ ⋮𝐵𝑖1 𝐵𝑖2 … 𝐵𝑖𝑁𝑠⋮ ⋮ ⋱ ⋮
𝐵𝑁𝑠1 𝐵𝑁𝑠2 … 𝐵𝑁𝑠𝑁𝑠]
,
𝐶 = [
𝐶11 0 … 00 𝐶22 ⋱ 0⋮ ⋮ ⋱ ⋮0 … … 𝐶𝑁𝑠𝑁𝑠
],
(2.68)
𝑢 = [𝑢1′ … 𝑢𝑁𝑠
′ ]′ ∈ ℝ𝑚,
𝑥 = [𝑥1′ … 𝑥𝑁𝑠
′ ]′ ∈ ℝ𝑛, (2.69)
MPC
Subsystem
1
Subsystem
2
u1
u2
System
y1
y2
x1 x2
LITERATURE REVIEW
37
𝑦 = [𝑦1′ … 𝑦𝑁𝑠
′ ]′ ∈ ℝ𝓏 .
For each subsystem i= 1, 2,…, Ns, the (ui, xi,, yi) represents the subsystem input, state and
output vector respectively. Thus, in the centralized MPC framework, the following optimization
problem is solved,
𝑚𝑖𝑛𝑥,𝑢
𝐽(𝒙, 𝒖; 𝑥(𝑘)) =∑𝑤𝑖𝑖
𝐽𝑖(𝒙𝑖 , 𝒖𝑖; 𝑥𝑖(𝑘)), (2.70)
𝑠. 𝑡. 𝑥(𝑘 + 1) = 𝐴𝑥(𝑘) + 𝐵𝑢(𝑘), (2.71)
𝑢𝑖(𝑘) ∈ 𝒰𝑖 𝑖 = 1, … , 𝑁𝑆,
𝐽𝑖 =∑𝐿𝑖
𝑁
𝑗=1
(𝑥𝑖(𝑘 + 𝑗), 𝑢𝑖(𝑘 + 𝑗 − 1)),
𝐿𝑖 =1
2[𝑥𝑖𝑇𝑄𝑖𝑥𝑖 + 𝑢𝑖
𝑇𝑅𝑖𝑢𝑖].
Where 𝑤𝑖 > 0,∑𝑤𝑖 = 1, 𝑄𝑖 ≥ 0 , 𝑅𝑖 > 0 and (𝐴𝑖, √𝑄𝑖) detectable and 𝑥𝑖(𝑘) given.
For any system, centralized MPC achieves the best attainable performance (Pareto optimal) as
the effect of interconnections among subsystems are accounted for exactly. Additionally,
existent conflicts among controller objectives are solved optimally (Venkat et al., 2008).
Normally, centralized MPC can be used in the form of a supervisory controller that have access
to all variables in the network and that determines actions for all actuators. Due to practical and
computational issues the implementing a centralized controller may not be feasible. Individual
hubs may not want to give access to their sensors and actuators to a centralized authority and
even if they would allow a centralized authority to take over control of their hubs, this
centralized authority could have computational problems with respect to required time when
solving the resulting centralized control problem. This approach is ideal for small scale systems,
but when the control and optimization problem involves a large scale system the conventional
approach requires the decomposition of the global system into smaller subsystems.
2.3.6 Decentralized MPC
Alternatively, decentralized control uses a distributed approach that decomposes a large
optimization problem in small ones that can be solve by agents, preserving or even improving
the final performance (Camponogara et al., 2002). Each agent solves its sub-problem and
computes its own control actions, taking in to account possible mutual influences between other
LITERATURE REVIEW
38
neighbouring agents. Thus, decentralized control is desirable for being scalable and robust but
fails to account for the mutual influence among neighbouring sub- systems that can have long-
range effects. The key feature of a decentralized control framework is that there is no
communication between the different local controllers. While there are some important reviews
on decentralized control see (Christofides et al., 2013) for a tutorial review. A schematic for
decentralized MPC architecture, with two subsystems with state coupling, is shown in Figure
2.13.
Figure 2.13. Decentralized MPC architecture (Scattolini, 2009).
Note that strong interactions between different subsystems may prevent one from achieving
stability and desired performance with decentralized control. So, in order to achieve closed-loop
stability as well as performance in the development of decentralized MPC algorithms, the
interconnections between different subsystems are assumed to be weak or negligible, and are
considered as disturbances which can be compensated through feedback so they are not
involved in the controller formulation explicitly. Thus, the effect of external subsystems on the
local subsystem is omitted in this modelling framework. Remark that, when the above
assumption is not valid it may lead to a reduced control performance. Assuming a block
diagonal structure for the overall dynamic system, each subsystem is modelled by
𝑥𝑖(𝑘 + 1) = 𝐴𝑖𝑖𝑥𝑖(𝑘) + 𝐵𝑖𝑖𝑢𝑖(𝑘), (2.72)
𝑦𝑖(𝑘) = 𝐶𝑖𝑖𝑥𝑖(𝑘) 𝑖 = 0, … , 𝑁𝑆.
In the decentralized MPC framework, each subsystem MPC solves the following optimization
problem,
𝑚𝑖𝑛𝑥,𝑢
𝐽𝑖(𝒙𝑖 , 𝒖𝑖; 𝑥𝑖(𝑘)), (2.73)
𝑠. 𝑡. 𝑥𝑖(𝑙 + 1|𝑘) = 𝐴𝑖𝑖𝑥𝑖(𝑙|𝑘) + 𝐵𝑖𝑖𝑢𝑖(𝑙|𝑘), 𝑘 ≤ 𝑙, (2.74)
𝑢𝑖 ∈ 𝒰𝑖 , 𝑘 ≤ 𝑙.
MPC 1
MPC 2
Subsystem
1
Subsystem
2
u1
u2
System
y1
y2
x2 x1
LITERATURE REVIEW
39
Each decentralized MPC solves an optimization problem to minimize its local cost function.
2.3.7 Distributed Model Predictive Control
The MPC have also evolved as a distributed systems control methodology (Negenborn, et al.,
2006; Scattolini, 2009; Igreja et al., 2012). Distributed or decentralised Model Predictive
Control allows the distribution of decision-making while handling constraints in a systematic
way. DMPC algorithms, offers the same features of MPC but have the advantage of supporting
the distribution of sensing and control using local controllers/agents that cooperate by
exchanging information to decide their control actions. This is the reason why it was the chosen
method to deal with this kind of system. These DMPC infrastructures are suitable to use in a
MAS (see Chapter 2.4) framework where, in a distributed environment, several agents
employing individually a MPC control strategy are able to interact and receive influence from
neighbour subsystems, exchange predictions on their future state and incorporate this
information into their local MPC problems.
Large scale interconnected systems, such as power systems, water distribution, traffic systems,
etc, require control techniques adjusted to the distributed nature of the system, as, decentralized
or distributed control (Krogh, 2001). However, in decentralized MPC approach, the information
exchange between subsystems is ignored and each subsystem is constructed with its individual
control system forming a traditional centralized control scheme. This decentralized control
architecture may result in poor system-wide control performance if the subsystems interact
significantly (Venkat, et al., 2006). In distributed MPC, there is not a single MPC controller but
instead there are multiple MPC controllers, each for a particular system. Typically, there is
dynamical interaction among the systems that the individual controllers consider. Each of the
controllers adopts the MPC strategy as outlined above for controlling its system but now not
only considering dynamics, constraints, objectives, and disturbances of the subsystem under
consideration but also considers the interactions among the systems. Each local controller solves
an MPC problem based on local information and may hereby share information with the other
controllers to improve the overall performance.
Figure 2.14 represent a simple distributed control system where it is assume that local regulators
interact between them, obtaining by this way some information on the behaviour of the others.
This information is normally related with future predicted control or state variables computed
locally, so that any local regulator can predict the interaction effects over the considered
prediction horizon.
LITERATURE REVIEW
40
Figure 2.14. Distributed MPC architecture (Scattolini, 2009).
The generalized optimal state-input trajectory for (𝒙𝑖𝑝, 𝒖𝑖𝑝) for subsystem i (i=1,2,…Ns) at
iteration p is obtained as the solution to the optimization problem,
𝑚𝑖𝑛𝒙𝒊,𝒖𝒊
𝐽𝑖(𝒙𝑖 , 𝒖𝑖; 𝑥𝑖(𝑘)), (2.75)
𝑠. 𝑡. 𝑥𝑖(𝑙 + 1|𝑘) = 𝐴𝑖𝑖𝑥𝑖(𝑙|𝑘) + 𝐵𝑖𝑖𝑢𝑖(𝑙|𝑘) +∑[𝐴𝑖𝑗𝑥𝑗𝑝−1(𝑙|𝑘) + 𝐵𝑖𝑗𝑢𝑗
𝑝−1(𝑙|𝑘)]
𝑗≠𝑖
, 𝑘 ≤ 𝑙 (2.76)
𝑢𝑖 ∈ 𝒰𝑖 , 𝑘 ≤ 𝑙, 𝑖 = 0, … , 𝑁𝑆.
When moving from a centralized MPC to a distributed MPC setting, several key concepts
become relevant. In a distributed MPC setting, a system or subsystem refers to the entity being
controlled by the controller. The overall system is the combination of all systems or subsystems
under control merged into one large-scale system. The notions of global versus local distinguish
between the overall system and the system or subsystem under control by a particular controller.
Hence, frequently appearing concepts are local objectives, local dynamics, local constraints, and
local disturbances. The terms interconnecting and shared are often used in combination with the
terms variables and constraints to denote explicitly those components that represent the
connections or couplings between different systems. In a distributed setting, a particular
controller has neighbours, or neighbouring controllers. The neighbours are those controllers that
control systems that are coupled or influence the system under control by this particular
controller. Communication takes places among the controllers, which can exchange
information, for example, regarding local states, local objectives, and/or local constraints.
Information can then be taken into account by the controllers to implement a coordination or
negotiation process. The controllers can be structured in control layers or levels, leading to a
hierarchical control structure. Here, typically at higher levels, controllers consider slower time
scales and larger systems in a more abstract way, whereas at lower levels, controllers consider
faster time scales and smaller systems in a more detailed way. The potential of distributed MPC
lies in the unique combination of the strengths of MPC (namely, feedback with feed-forward
MPC 1
MPC 2
Subsystem
1
Subsystem
2
u1
u2
y1
y2
System
x2 x1
LITERATURE REVIEW
41
control in a receding horizon fashion, multi-objective optimization, and explicitly handling of
constraints) with the negotiation and coordination possibilities provided by communication. The
role played by communication in this context is essential. For example, the distinction between
decentralized and distributed MPC lies in whether or not the controllers actively communicate
with one another to determine actions. In decentralized control architecture, there is no direct
communication between controllers; controllers take into account the influence of neighbouring
systems only by responding to the dynamics of the systems they are controlling. This constraint
typically limits the performance that the decentralized MPC control scheme can achieve. In fact,
in a decentralized MPC scheme, controllers may be opposing one another’s actions, even
though they may not have the intention to do so. In a distributed MPC setting, such a situation
can be prevented because the controllers can communicate about what actions they are going to
take when and obtain agreement on an optimal timing.
In summary, DMPC is an evolving technology that advocates the distribution of sensing and
control while preserving the same features of standard MPC, namely, the use of a prediction
model and the explicit handling of constraints (Saluje et al., 2011).
As mentioned DMPC strategies can be characterized by the type of couplings or interactions
assumed between constituent subsystems (Trodden & Richards 2010), (Keviczky, et al., 2006).
For example, dynamically coupled systems can be seen in several contributions (Doan, et al.,
2009; Camponogara et al. 2002; Ling, et al., 2005; Dunbar 2007; Giovanini, et al., 2007;
Venkat, et al., 2008, Alessio, et al., 2011).
Another common application of DMPC is on subsystems where the states are not dynamically
coupled but are coupled in their objective functions. A DMPC coupling via the cost function
(Wang, et al., 2010), considers the DMPC of systems with interacting subsystems having
decoupled dynamics and constraints but coupled costs. Works with similar approaches can be
seen in literature (Raffard, et al., 2004; Franco, et al. 2007).
Other different strategy involves subsystems sharing coupled constraints and many approaches
have been presented in the literature. In Riggs, et al., 2010 is establish an algorithm where an
optimization problem is performed by different subsystems with their own objective functions
and local constraints, but share a common coupled constraint which limits the behaviour of the
independent systems.
Biegel, et al.,(2014) presents a dual decomposition as a means to coordinate a number of
subsystems coupled by state and input constraints where each subsystem have a model
predictive controller while, a centralized entity manages the subsystems via prices associated
with the coupling constraints. This system allows coordination of all the subsystems without the
need of sharing local dynamics, objectives and constraints.
LITERATURE REVIEW
42
Another approach (Müller, et al., 2012) presents a framework for DMPC of discrete-time
nonlinear systems with decoupled dynamics, but subject to coupled constraints and a common,
cooperative task. Each system exchange information about is predicted trajectories with its
neighbours in order to be able to achieve the common objective and to satisfy the coupling
constraints. Systems coupled by constraints can be seen in (Waslander, et al., 2004; Keviczky,
et al., 2006; Richards & How 2007; Kuwata et al., 2007).
Due to its features, DMPC (Camponogara, et al., 2002) has been developed for application to
large-scale systems, namely as process control (Borrelli, et al., 2005), chemical plants (Venkat,
et al., 2004) and or teams of vehicles (Riggs, et al., 2010) (Kuwata, et al., 2007), in which
control by a single centralised agent would require excessive communication, computation and
reliance on a single processor.
Due the characteristics mentioned above DMPC, are been applied in many distributed
applications, and several approaches and architectures to this control scheme are emerging.
Some studies are applying MPC to reduce and optimize the energy consumption in the
residential sector, namely, to deal with temperature set points regulations (Morosan, et al., 2010;
Bălan, et al., 2009; Freire, et al., 2008). In particular Freire, et al., (2008), with a DMPC control
algorithm based a modified version of Benders’ decomposition the control problem objective is
to minimize the heating energy bills while maintaining a certain indoor thermal comfort.
A DMPC scheme with stability constraint (DMPC-SC), where controllers to obtain coordination
with each other, spread information about their predictions and their local state measurements
and use them in their local computations, is described in (Krogh, 2001).
For serially connected processes, (Zhang & Li, 2007), proposed a networked MPC scheme with
neighbourhood optimization where each MPC unit receives information about the foreseeable
action for its neighbour sub-processes. Therefore, the exchange information allows cooperation
between subsystems and the optimizer to generate a suitable control decision for each sub-
process.
Bendtsen et al., (2010), proposed hierarchical approach, consisting of a three-level structure
with a high level MPC controller, a second level of so-called aggregators, controlled by an
online MPC-like algorithm, and a lower level of autonomous units. The approach is motivated
by smart-grid electric power production and consumption systems, being the goal to
accommodate load variations on the grid, arising from varying consumption and natural
variations in power production, e.g. from wind turbines.
Stewart et al., (2010) also proposed a hierchical cooperative distributed MPC without requiring
additional coordinating controllers. The head of each subsystem performs a small additional
LITERATURE REVIEW
43
calculation at each pack of information exchange to reduce the volume of communication and to
hide input trajectories between neighbours.
Negenborn, (2007), develop a multi-agent MPC control structure with applications to control
problems in power networks. The control agent uses the prediction model to predict the
behaviour of the system under various manners, and establish the actions that optimize the
behaviour of the system and minimize the costs existing in the objective function.
Ventak (2006), proposed, a cooperative distributed MPC framework, where the objective
functions of the local MPCs are modified to achieve global control objectives.
Besides the advantages above mentioned, DMPC seems to be the right approach to deal with
complex distributed systems, they can take into account all available information and that it can
therefore anticipate undesirable situations in the future at an early stage.
As mentioned, for the control problem of large-scale systems, centralized model predictive
control is impractical due to a number of limitations, including the computational efforts and
limited communication, and due this fact, more research is needed in distributed and
hierarchical model predictive control methods in order to develop new methods and algorithms
to implement in large-scale systems. Smart grids are characterized by complex dynamics and
mutual influences between all the involved identities and the algorithms should be able to deal
with couplings in the dynamics and the constraints between subsystems, and guarantee
feasibility and stability of the closed-loop system.
Instead, DMPC distributes control decision-making among agents corresponding to the different
subsystems making up the whole. The challenge is then how to coordinate efforts to ensure that
the distributed decisions lead to constraint satisfaction, feasibility and stability of the overall
closed-loop system. Several strategies for DMPC have been presented in the literature, and
many theoretical results exist, including those for feasibility and stability; see Scattolini (2009)
for a comprehensive survey. The approaches are broadly divisible by the type of couplings or
interactions assumed between constituent subsystems. The degree of coupling among the
subsystems varies. In the most complex situation, the dynamics or/and constraints of
subsystems are coupled. The method presented assumes the latter type of coupling, and has
agents update their plans one at a time, with iteration, to ensure coupled constraint satisfaction;
however, unlike other methods, it also permits a flexible order of updating.
Robustness to disturbances is a key challenge in the development of MPC (Mayne et al., 2000),
and is harder still when control decision-making is decentralised; few DMPC schemes in the
literature offer robustness. In Richards and How (2007), robust feasibility and stability are
guaranteed by updating each subsystem’s plan in a sequence, subject to tightened constraints,
and while ‘freezing’ the plans of others. Alternative approaches include treatment of
LITERATURE REVIEW
44
interconnected subsystems’ state trajectories as bounded uncertainties, and using min-max
optimization (Jia & Krogh 2002) – though the complexity issues with such an optimisation
method are well documented (Mayne et al., 2000). Using the comparison model approach to
robustness (Fukushima & Bitmead 2005), another distributed method (Kim & Sugie 2005) uses
worst-case predictions of state errors, determined based on a robust control Lyapunov function,
and tightens constraints accordingly. Magni & Scattolini (2006) propose a robust stable
decentralised algorithm for non-linear dynamically coupled systems, with no information
exchange between agents, although for an asymptotically decaying disturbance.
The new algorithm in this thesis exploits this feature to achieve flexibility in communication.
An additional advantage of this approach is that the optimisation involves only the nominal
system dynamics, avoiding the large increase in computational complexity associated with the
inclusion of uncertainty in the optimisation as mentioned in (Scokaert & Mayne 1998). Many
distributed methods proposed in the literature (e.g. Du et al., (2001), (Kim & Sugie 2005),
(Dunbar & Murray 2006), (Alessio & Bemporad 2007), (Richards & How 2007), (Venkat et al.,
2008) do not consider the implications that the scheduling of local optimisations has on the time
required for communications. For example, the constraint-tightening DMPC approach proposed
by (Richards & How, 2007), also for dynamically decoupled systems with coupled constraints,
assumes repeated instantaneous exchanges during each sampling period.
2.4 Multi-Agents systems (MAS)
There are several ways to define an agent. Despite all the definitions, the most common view
describes agents like been a high-level software abstraction, which efficiently describes a
complex software entity, or, intelligent entities with three main characteristics:
Pro-Activeness (they react to external events and they are driven to their objectives);
Social ability (they can cooperate or compete between them);
Autonomy (they can decide in order to archive their objectives);
Multi-agent systems can be seen as a group of distributed, autonomous and intelligent
agents within an environment, where they can act and react in order to work together to
achieve a global goal.
Large interconnected networks can hardly be controlled in a traditional way due to its
distributed characteristics and considerable control capabilities over the network operation.
Multi-agent systems have characteristics that meet these requirements, they are able to model
complex systems introducing the possibility of agents having common or conflicting goals.
They may decide cooperating for mutual benefit or may compete to serve their own interests
(Xu et al., 2010). Thus, from the single agent concept to the group of agent’s concept, MAS are
LITERATURE REVIEW
45
a powerful means to represent systems in a natural way. A system is traditionally defined as a
group of components interacting together to achieve an overall objective or a number of
objectives known as the system level objective(s). Each component in a system can be seen as
an agent if it possesses the characteristic of agency, i.e. if it performs computations and its
action is being feedback into the environment (Abbass et al., 2011). Distributed infrastructures
are interconnected and maintain communication between each other is suitable for MAS
technology (Raza et al., 2005) due the follow characteristics:
Agents assess situation on the basis of measurements from sensing devices or that
received from other entities;
Agents cast their influence on performance of networks via commands to actuators and
other identities;
The agents differ in complexity from trivial threshold detectors that merely decide
based upon a single measurement to highly intelligent systems;
Agents have to make real time decisions from local, rather than global state
information;
Agents including controllers for individual devices are designed with simple decision
rules based upon response thresholds that are expected to give most appropriate
responses to a collection of situations generated.
Smart grids can be seen as an environment where a set of simple entities called agents can play
the role of source, load, or any other component integrated in the environment. With the agents
interaction, some collective intelligent can be achieved, they cooperate or compete in other to
reach their objective.
MAS architectures have also others advantages, they are not dependent on a particular
technology, with proper messaging tools, different programming languages can be used and
interact together (Roche et al., 2010), and, due the agents autonomy, they can leave or enter in
the environment without any perturbation.
2.4.1 MAS architectures
Researchers are still attempting to categorize the different types of MAS and agents within
MAS. They are trying to identify characteristics and features that an agent must possess and also
by defining categories of agents and MAS. In (Ferber, 1999) is presented a number of
categorizations of MAS and one of the most accepted formal definitions of as: MAS= {E, O, A,
R, Or, Op} where:
LITERATURE REVIEW
46
E is the environment, a space in which objects (including objects O and agents A) are
located;
O is a set of objects in the environment other than the agents;
A is the set of agents in the environment;
R is an assembly of relations;
Or is an assembly of operators; and,
Op is an assembly of operations.
Due to the flexibility of MAS, several control architectures can be also found in order to
respond to the variety of power systems, such as, voltage control, microgrid management,
distributed network control and restoration, to manage and obtain the desire degree of
intelligence. However, some general types of control structures can be identified and they are
commonly encountered in theory and practice, single-agent, multi-agent single-layer and multi-
layer. A single-agent control structure,
Figure 2.15 occurs when it is assumed that there is only one control agent that has access to all
actuators and sensors of the network and thus directly controls the physical network. This
control structure is referred to as an ideal, since in principle such a control structure can
determine actions that give optimal performance.
Control Agent
Measurements Actions
Control structure
Figure 2.15. Single-agent control structure. (Adapted from Negenborn 2007).
When there are multiple control agents, each of them considering only its own part of the
network and being able to access only sensors and actuators in that particular part of the
network, then the control structure is referred to as a multi-agent single-layer control structure,
Figure 2.16. If in addition the agents in the control structure do not communicate with each
other, the control structure is decentralized. If the agents do communicate with each other (dot
line), the control structure is distributed.
LITERATURE REVIEW
47
Control Agent 1
Measurements Actions
Control Agent 3
Actions Measurements
Control Agent 2
Measurements Actions
Control structure
Figure 2.16. Multi-agent single layer control structure. (Adapted from Negenborn 2007).
A multi-layer control structure can be defined when there are multiple control agents and some
of these control agents have authority over other control agents, in the sense that they can force
or direct other control agents,
Figure 2.17. A multi-layer control structure typically is present when one control agent
determines set-points to a group of other control agents working in a decentralized or distributed
way. Due to the authority relationship between agents or groups of agents, this structure can
also be referred to as a supervisory control structure, or a hierarchical control structure.
Control Agent 1
Measurements Actions
Control Agent 3
Actions Measurements
Control Agent 2
Measurements Actions
Control Agent 4
Control structure
ActionsMeasurements
Figure 2.17. Multi-layer control structure. (Adapted from Negenborn 2007).
Single-agent control structures in general, have the advantage of deliver the best performance
possible, and that they have been studied extensively in the literature, in particular for small-
scale systems. Although, in large scale systems, there are several issues related with, robustness,
reliability, scalability, responsiveness and communication delays that complicate the use of this
structure.
Multi-agent control structures can deal or at least reduce these issues, but typically, they have a
lower performance comparatively with single-agent.
Decentralized multi-agent single-layer control structures have the advantage, over the
distributed, of lower computational requirements and faster control, because the inexistence of
LITERATURE REVIEW
48
communication between the controllers. However, this advantage will typically be at the price
of decreased overall performance. The advantage of a distributed multi-agent single layer
control structure is therefore that improved performance can be obtained, although at the price
of increased computation time due to cooperation, communication, and perhaps negotiation
among control agents. The multi-agent multi-layer control structure, despite of been more
complex, provides the possibility to obtain a trade-off between system performance and
computational complexity (Negenborn 2007).
Although all the existing diversity, to respond to the several challenges mentioned above, most
of the authors use a hierarchical layered structure (Zeng et al., 2009; Dimeas et al., 2005; Li et
al., 2010; Aung et al., 2008) to archive different objectives.
With the joint goal of achieving cost and energy savings Ab et al., (1996), shows how
decentralized power load management at the customer side, automatically carried out by a
‘society’ of intelligent household, industrial and utility equipment, can be modelled in terms of
independent hierarchical decentralized intelligent agents that communicate and negotiate in a
computational market economy.
Lu & Chen (2009) proposes new a system-level scheme, a three layered MAS architecture,
where the bottom layer is called physical layer, in which each agent correspond to one energy
source (RE or traditionally), the middle layer is named application layer, which is composed of
coordinator agent, fault diagnosis agent, and recover agent etc., which provides service to the
physical layer. The top layer is the user interface layer, which enables the interaction between
human and MAS.
Also with a hierarchical layered structure, the MAS main target present by Jian et al., (2009), is
to maximize the efficiency of renewable energy resources.
Pipattanasomporn et al., (2009), describes the design and implementation of the multi-agent
system for use in a microgrid. The proposed multi-agent system consists of a control agent, a
DER agent and a user agent in multi-layer distributed architecture where all agents interact
between them. The proposed design also includes a database agent that is responsible for storing
system information, as well as recording the messages and data shared among agents.
In order to negotiate available energy quantities and needs on behalf of consumers and producer
groups, Wedde et al., (2006 and 2008) developed a multi-level bottom-up solution with
autonomous software agents.
Negenborn, (2007), develop a multi-agent control structure with applications to control
problems in power networks, and compare the obtained results with the single-agent approach.
LITERATURE REVIEW
49
As showed, to respond to the several issues related with distributed nature of the future smart
grids, the MAS structure demonstrate to be the best solution. The existent MAS designs mostly
use layered and subsystems models were a group of agents is controlled by a superior control
agent. However, large distributed systems may have innumerable identities, and due the
necessary communication between agents, it is needed a high computational effort. Besides this
fact, the increase of agents also increases the possibility of internal and external conflicts,
leading to a possible performance reduction.
In order to solve these problems, it is necessary to develop new algorithms, control system
designs, and MAS architectures that can provide the benefits of cooperation, communication
and negotiation among control agents to increase the performance, but with lowest
computational efforts, and also to reduce the complexities to obtain an optimal operation in a
large-scale distributed environment.
LITERATURE REVIEW
50
51
Chapter 3
3 Dynamical Models and Scenarios
Description
3.1 Introduction
The energy consumption optimization in future SGs will be based on grid-integrated near-real-
time communications between various grid elements in generation, transmission, distribution
and loads.
Figure 3.1. Typical smart grid architecture.
SELECTION AND IDENTIFICATION SCENARIOS AND MODELS
52
Figure 3.1 shows a typical smart metering architecture that is being reflected in the European
standards development process (Fan et al., 2013).
Smart Grid is a concept for transforming the electric power grid using advanced automatic
control and communications techniques and other forms of information technology. It integrates
innovative tools and technologies from generation, transmission and distribution all the way to
consumer appliances and equipment. This concept integrates energy infrastructure, processes,
devices, information and markets into a coordinated and collaborative process that allows
energy to be generated, distributed and consumed more effectively and efficiently. Costumers
will have the chance to decrease their electricity bills by changing consumption from higher-
priced hours to lower priced hours and smart metering will produce a market for new SG
costumer products.
Thus, the scenario here described intends to be a realistic solution to take advantage of these
SG’s features.
3.2 Scenarios overview
A SG generally involves the application of smart meters, sometimes called advanced metering
infrastructure (AMI), which usually include control and monitoring devices and appliances.
Smart metering technology will be at the foundation of any SG design (Cecati et al., 2010). The
most viable communication’s technologies for the AMI are wireless and power line
communication (PLC). SG’s communication’s infrastructure will also provide IP/Ethernet
connectivity between most components, guiding the communication’s interfaces and products
towards TCP/IP-based networks. Power distribution companies are mostly interested in PLC
because it represents the most cost effective solution and does not require additional investment
in the communication’s infrastructure. For these communication’s requirements, companies like
Siemens are offering customized communications network solutions for optic fibre, power line,
and wireless infrastructures based on the accepted standards of the energy industry, Figure 3.2.
Figure 3.2. Typical energy distribution communication architecture for smart grids (Adapted
from Siemens, 2014).
SELECTION AND IDENTIFICATION SCENARIOS AND MODELS
53
Because renewable energies are nowadays significantly expanded, electricity is being fed into
both the medium- voltage and low-voltage grids, depending on changing external conditions
(e.g., weather, time of day, etc.). These fluctuating energy resources can severely impair the
stability of the distribution grid. One of the key challenges of a smart grid is therefore to quick
balance out the energy supply and energy consumption in the distribution grid.
Thus, the scenario here presented has in consideration this challenge, intend to provide a
solution to it based on, and taking advantage of all the AMI and communication technologies
that SG provide.
The conceptual scenario involves a set of buildings with an electricity source provided by their
own renewable energy park and energy storage as presented in Figure 3.3. Henceforward, the
term house will be applied to classify any type of structure for habitation (houses), office
buildings or other kind of analogous constructions. The set 𝑊 = {𝑤1, 𝑤2, … , 𝑤𝑁𝑆} defines the
Thermal Control Areas TCA’s/house and the different spaces are described by 𝐷ℎ =
{𝑑h1, 𝑑h2, … , 𝑑ℎ𝑁𝑑} and h = 1,…𝑁𝑆 .
Figure 3.3. Implemented scheme.
Due the existing division diversity in houses, each area may vary in: construction materials, sun
exposure, occupancy, indoor temperature set-points. This variety involves many different
energy needs to weatherize the spaces, and for this a TCA is considered. Figure 3.4 presents a
conceptual TCA smart thermostat which can be used as an interface for DSM. In the left side of
the display the user is able to choose between three operating modes related with the control
features. At the centre the user is able to program the division’s comfort range and the particular
time periods. On the right the cost, the green and red consumption are shown, allowing the user
to access more specific other menus.
SELECTION AND IDENTIFICATION SCENARIOS AND MODELS
54
Figure 3.4. Thermal Control Area (TCA) smart thermostat controller vision.
Each division may have different thermal loads, thermal characteristics, occupancy and comfort
temperature bounds and consequent different energy needs for heating/cooling the spaces, for
this a Thermal Control Area (TCA) is considered. As mentioned, one TCA represents an
autonomous thermal control entity within an environment where the actions and reactions are
made in order to achieve a common goal. Therefore, as Figure 3.5 shows, depending from the
desired infrastructure intent to be implemented, a set of buildings or a simple division may
represent a TCA.
Figure 3.5. Thermal Control Area concept.
The idea is to use a MPC control law to adjust the indoor temperature set-point in each TCA.
Taking advantage of the MPC characteristics, the system receives information on the outdoor’s
temperature forecasts, the future occupation, the indoor temperature frame and the thermal
disturbances profile. The control action is conditioned by two constraints; thermal comfort and
available energy. Using a DSM approach, the trade-off between the energy price and comfort
limitations are system assured. The scenario considers the existence of two resources, the red,
from the grid provided by fossil source and the green, from renewable or clean source. The first
TEMP. MAX:
TEMP. MIN:
ECO MODE
COMFORT
NORMAL21º
19º
SET TIME PERIOD
13 15-
13:30
SETTINGS
COST
2.5
COMSUNPTION
2.5 KWH
GREEN
0.7RED
78%
22%
INDOOR TEMP.
20º €
SELECTION AND IDENTIFICATION SCENARIOS AND MODELS
55
is always available with a kWh price always higher than the green. On the contrary the green is
limited, intermittent and must be consumed, stored or grid delivered. Remark that if the clean
source becomes insufficient the fossil is consumed increasing the energy costs.
So, the scenario considers that the traded electricity comes from a wholesale market which is
priced in real time, where the market operators (MO) provide the green resource auction and the
agents/TCA set a day-ahead their bids. The agents (one by each TCA) must bid in an auction,
the price that they are willing to pay to consume the green resource. The bid value establishes a
sequential order by which the TCA’s can access to green energy. After consuming, each TCA
sends to the next the information about how much of clean resource is still available. As
mentioned, when the green resource becomes insufficient to satisfy all the demand, the red is
available. The red resource consumption implies a penalty in the final cost function (5.1) due to
the soft constraint violation imposed by the maximum available green resource. Notice that the
divisions that interact thermally pass to each other information about future temperatures.
Each TCA receives several external inputs; the outdoor temperature, available renewable power
forecasts, the MO kWh price, access order to the renewable resource and, when applied, the
neighbour’s indoor temperature forecasts. Figure 3.6 presents an example of the implemented
TCA framework..
Figure 3.6. Example of a TCA conceptual framework.
The example depicted in Figure 3.6 shows the implemented system dependent from weather and
power forecasts. Being wind and solar power not dispatched the production levels need to be
predictable. Lower prediction errors help to balance the supply of wind energy with other
sources, avoid consequences for incorrect estimation, as well as reducing the risk of having to
buy energy from elsewhere. For wind power, the 24 hours ahead prediction average absolute
error can be between 10–20% of installed capacity for one single zone. When aggregation of
Market Operator/
Energy Supplier
Wheather
Forecasts
Available Renewable
Power Forecasts
Neighbors indoor
temperature forecast
SELECTION AND IDENTIFICATION SCENARIOS AND MODELS
56
zones is made, the value may drop to 5-8% (VTT Technology 95, 2014; NREL, 2014). This can
be partly due to the different accuracy for different types of sites, like different orography and
roughness and how homogenous the grid cell for the numerical weather prediction model is.
Prediction errors in low wind situations can still be higher relative to yearly energy. The forecast
error is significantly lowered for the first six hours. The physical reason behind this is that wind
itself has an auto-correlated nature and the model takes this into account by last measured power
value when making forecasts into the future. As the forecast horizon increases the variation
range seems to increase quite linearly.
The forecasting of solar energy production faces issues similar to those of wind. However, solar
forecasting has significant predictability because the sun’s path through the sky is known.
Accurate outdoor temperature forecasts are particularly valuable for electric or gas utilities.
Nowadays, temperature forecasts are particularly reliable and accurate, with more than 92%
accuracy with one day ahead (ForecastWacth, 2014).
Thus, in a scenario that privileges the renewable energy usage, the used demand side
management approach allows the management of distributed loads, aiming the adjustment of the
demand to the supply, providing thermal comfort, lower energy costs and lowering
CO2.emissons. By using this active DSM control, the optimal control strategies for various
appliances can be generated whilst maximum utilisation of energy supplied from intermittent
systems is guaranteed. The used DSM strategy is described in Chapter 3.3.
3.3 Demand side management approach
It is widely accepted that some form of real-time pricing arrangements is required to efficiently
allocate DSM resources and fully inform users about the value of electricity at each point in
time and location. Combination of ICT and business processes will be significant tools in the
real time management of active networks, customers and commercial systems. Deregulated
markets will let consumers to exploit information to move between competing energy suppliers
based on energy cost, greenhouse gas emissions and social goals. A SG vision for the market
has been developed in recent years through the work of a consortium of utilities. One option is
an “eBay for electricity”, where continual electronic sales match energy consumers with energy
producers. Soon customers will be able to bond the electricity pricing into their energy
management system. High gap in price differentials between cost periods could result in greater
shifts of energy usage. This needs to be accompanied by the application of intelligent appliances
that would facilitate the implementation of DSM. In order to simplify such trading of energy
among a very large number of smaller domestic participants, an electronic energy market
SELECTION AND IDENTIFICATION SCENARIOS AND MODELS
57
system, supported for example by the internet, would need to be developed (an extension of
power-exchange-type markets).
Thus, the developed system intends to provide a solution to be implemented in these future
energy markets. It considers that the MO offers a green resource auction where each TCA sets a
day-ahead their bids. The agents (one by each TCA) must bid in an auction, the price that they
are willing to pay to consume the green resource according to their own consumption forecasts.
Utilities are able to see what power loads are creating bigger demand on their side, allowing
them to more effectively implement DR programs for heating and cooling. Although, before
demand–response systems can be effectively deployed on a wide scale in the residential sector,
a number of technical challenges need to be resolved (infrastructure of communications,
metering infrastructure, demand–response thermostats, etc.). This is likely to include some form
of house energy management system that may rely on wireless technology that would
automatically respond to price signals while taking into consideration the homeowner's
preferences for cost versus comfort. Furthermore, the thermostat can be programmed to change
settings with seasons. The thermostats can also have a notification feature to alert residents for
calls for action, as well as an override feature, in case the customer chooses not to participate in
the particular event. The customer main obtain information on buy-back rates via internet
connections and takes appropriate action to manage peak loads. A key issue in these
programmes is how sophisticated or complex they need to be to make the price signals. There is
also the issue of verification to confirm that some benefit was obtained when the thermostat and
air-conditioning system responded. Thermostats can also be pre-programmed to receive weather
forecasts, to optimize both a consumer's energy savings and grid performance. The system uses
weather forecasts and, when applied, neighbours indoor temperature forecasts, to anticipate
major changes in the surrounding temperature and manage heating and cooling needs in
advance, in the most energy-efficient way.
3.4 TCA Dynamical Models
The dynamics of temperature evolution in a building is one of the most important aspects of the
overall building dynamics. The complexity in the dynamics of temperature evolution comes
from the thermal interaction among rooms (and the outside). This interaction can either be
through conduction across the walls, or through convective air exchanging among rooms.
Consumption of energy is predominantly determined from a selection of materials and
architectural solutions. It can be further reduced with efficient management of heating or
cooling.
SELECTION AND IDENTIFICATION SCENARIOS AND MODELS
58
The building geometry constitutes the basic input for energy simulation. It is crucial to
understand the differences between a building model created by an architect and a building
model needed for energy simulation. For example, energy simulations spaces need to be defined
by space boundaries, which are not necessarily the same as walls in an architectural model.
Thermal building simulation engines like, EnergyPlus or DOE-2 (Maile et al., 2007), predict the
energy performance of a given building and thermal comfort for its occupants. In general, they
support the understanding of how a given building operates according to certain criteria and
enable comparisons of different design alternatives. This kind of software allows designing and
characterizing in detail all the major features of a building architecture. Most of them also
require a set of inputs which mainly consist in building geometry, internal loads, HVAC
systems and components, weather data, operating strategies and schedules and simulation of
specific parameters. However, these tools are only used for thermal simulation with predicted
scenarios and previously established data, not working with real time inputs and data, which
allow instant action over devices with control and decision mechanisms.
Figure 3.7. Energy modelling and building design process example with EnergyPlus.
Every energy simulation is based on thermodynamic equations, principles and assumptions.
Since thermal processes in buildings are today complex and not totally understood, energy
simulation programs estimate their predictions with qualified equations and methods. Therefore,
results can be arbitrarily incorrect, if certain assumptions are not satisfied or matched in the
simulation in real life. While two or three dimensional heat transfer would increase the accuracy
of simulation results, geometry input and simulation running at real time is likely to become
more complex. As mentioned earlier, the input, especially weather data and internal loads, for
energy simulation is already based on assumptions, as are the thermodynamic concepts. For that
SELECTION AND IDENTIFICATION SCENARIOS AND MODELS
59
matter, any simulation is based on assumptions; therefore complex interrelations can be
simplified and managed. Users need to be aware of these assumptions and be able to decide
whether they are reasonable for their specific simulation or not. Thus, the idea here presented is
to apply the principle of analogy between two different physical domains that can be described
by the same mathematical equations. Consequently, a linear electrical circuit represents the
building, the state-space equations are obtained by solving the circuit. Here, the temperature is
equivalent to voltage and the heat flux to current. Heat transmission resistance is represented by
electrical resistance and thermal capacity by electrical capacity. The building equivalent circuit
is obtained by assembling models of the walls, windows, internal mass, etc. For single-zone
buildings, interior walls are part of the internal thermal mass, while exterior walls form the
building envelope. Several approaches are seen in (Zong et al., 2012; Hazyuk et al. 2011) where
is shown that building models can be simple or more complex depending on the objective.
Therefore, the first order energy balance model (3.1)-(3.3) that describes the dominant dynamics
of a generic division, is considered suitable for control purposes (Barata et al., 2014b). Note that
these are basic equations for edifice thermal modelling which can be described by several
divisions and floors thermally interacting between them. For complete thermal models see
(Hazyuk et al. 2011).
𝑑𝑇𝑙𝑖
𝑑𝑡=
1
𝐶𝑙𝑒𝑞𝑖(𝑄𝑙
𝑖ℎ𝑒𝑎𝑡
− 𝑄𝑙𝑙𝑜𝑠𝑠𝑒𝑠𝑖 + 𝑄𝑙𝑃𝑑
𝑖 ), (3.1)
𝑄𝑙𝑙𝑜𝑠𝑠𝑒𝑠𝑖 =
𝑇𝑜𝑢𝑡 − 𝑇𝑙𝑖
𝑅𝑙𝑒𝑞𝑖
+ ∑𝑇𝑔𝑖 − 𝑇𝑙
𝑖
𝑅𝑙𝑔𝑒𝑞𝑖 𝐶𝑙𝑒𝑞
𝑖𝛥𝑡
𝑁𝑑𝑖
𝑔=1(𝑔≠𝑙)
+ ∑ ∑𝑇𝑚ℎ − 𝑇𝑙
𝑖
𝑅𝑙𝑚𝑒𝑞𝑖ℎ 𝐶𝑙𝑒𝑞
𝑖
𝑁𝑑ℎ
𝑚=1
𝑁𝑠
ℎ=1(ℎ≠𝑖)
, (3.2)
𝑅𝑙𝑒𝑞𝑖 = 𝑅𝑙𝑟𝑜𝑜𝑓
𝑖 //𝑅𝑙𝑤𝑖𝑛𝑑𝑜𝑤𝑠𝑖 //𝑅𝑙𝑤𝑎𝑙𝑙𝑠
𝑖 //𝑅𝑙𝑡ℎ𝑖 , (3.3)
𝑅𝑙𝑤𝑖𝑛𝑑𝑜𝑤𝑠𝑖 =∑𝑅𝑤𝑖𝑛𝑑𝑜𝑤𝑚𝑎𝑡𝑒𝑟𝑖𝑎𝑙𝑠 , (3.3a)
𝑅𝑙𝑟𝑜𝑜𝑓𝑖 =∑𝑅𝑟𝑜𝑜𝑓𝑚𝑎𝑡𝑒𝑟𝑖𝑎𝑙𝑠 , (3.3b)
𝑅𝑙𝑤𝑎𝑙𝑙𝑠𝑖 =∑𝑅𝑤𝑎𝑙𝑙𝑚𝑎𝑡𝑒𝑟𝑖𝑎𝑙𝑠 , (3.3c)
where in (3.1), 𝑄𝑙𝑙𝑜𝑠𝑠𝑒𝑠𝑖 is heat and cooling losses (kW) from division (l), 𝑇𝑙
𝑖 the inside
temperature (ºC), 𝐶𝑙𝑒𝑞𝑖 the equivalent thermal capacitance (kJ/ºC), and 𝑄𝑙
𝑖ℎ𝑒𝑎𝑡
the heat and
cooling power (kW) and 𝑄𝑙𝑃𝑑𝑖 the external thermal disturbances (kW) (e.g. load generated by
occupants, direct sunlight, electrical devices or doors and windows aperture to recycle the
SELECTION AND IDENTIFICATION SCENARIOS AND MODELS
60
indoor air). In (3.2) 𝑇out is the outdoor temperature (ºC), 𝑅𝑙𝑔𝑒𝑞𝑖 the thermal resistance between
division (l) and the adjacent zone (g), 𝑅𝑙𝑒𝑞𝑖 the equivalent thermal resistance (C/kW), and 𝑅𝑙𝑡ℎ
𝑖 the
air thermal resistance to bulk of division.
Figure 3.8 and Figure 3.9 present a simplified schematic representation of thermal-electrical
modular analogy described by (3.1)-(3.3). 𝑇1𝑖…𝑇𝑁𝑑𝑖
𝑖 are the inside temperature of adjacent
spaces inside the same house, and 𝑅1𝑖 …𝑅𝑁𝑑𝑖
𝑖 the equivalent thermal resistance between the
division and that adjacent areas.
Figure 3.8. Schematic representation of thermal-electrical modular analogy for one division.
Figure 3.9. Generic schematic representation of thermal-electrical modular analogy for several
divisions (Barata et al., 2014b).
Tout
...
Ceq
Req
heat
i
l Qu
i
l
i
l Tx
...
Pd
i
l Qv
ii xT 11
i
Nd
i
Nd iixT
iR1
i
Nd iR
Thermal interconnections
with division of neighboring
houses
... Thermal interconnections
within the same house
T1Tout
Qheat1 Qlosses1C1
Req1
T2Tout
Qheat2 Qlosses2C2
Req2
TnTout
Qheatn QlossesnCn
Reqn
R2n
R12
R1n
...
...
T2, C2, Qheat2,
Req2, , QPd2
Tout
T3, C3, Qheat3,
Req3, QPd3
T1, C1,
Qheat1, Req1
QPd1
SELECTION AND IDENTIFICATION SCENARIOS AND MODELS
61
As described in (Barata, et al., 2013b), a first order energy balance model is used to define the
dominant dynamics of a generic division. Remark that these are the basic equations for structure
thermal modelling and can be described by several divisions and floors interacting between
them. In literature there are several degrees of complexity for house modelling techniques
(Thavlov, 2008). The model used is linear and considers control purposes suitability,
𝒙(𝑘 + 1) = 𝑨𝒙(𝑘) + 𝑩𝒖(𝑘) + 𝒗(𝑘), (3.4)
where x ℝn is the state variable, indoor temperatures, vector containing all the divisions
temperatures (ºC), u ℝm is the input vector containing all the heating and cooling power
sources (W) to weatherize each division, and v ℝn includes all the disturbances (W),
including load generated by occupants, solar radiation, any other heat/cooling sources or doors
and windows aperture to recycle the indoor air, k is an integer number that denotes discrete time
and A ℝn×n, B ℝn×m, are matrices.
Generically, using Euler discretization (Bequette, 2003) with a sampling time of , the discrete
model space-state representation of (3.4) can be written, for the Ns TCAs and for each division
(l) as,
𝑇𝑙𝑖(𝑘 + 1) = 𝐴𝑙𝑙
𝑖𝑖𝑇𝑙𝑖(𝑘) + 𝐵𝑙
𝑖𝑢𝑙𝑖(𝑘) + ∑ (𝐴𝑙𝑔
𝑖𝑖 𝑇𝑔𝑖(𝑘))
𝑁𝑑𝑖
𝑔=1(𝑔≠𝑙)⏟
𝑡ℎ𝑒𝑟𝑚𝑎𝑙 𝑐𝑜𝑛𝑡𝑟𝑖𝑏𝑢𝑡𝑖𝑜𝑛𝑠 𝑓𝑟𝑜𝑚 𝑎𝑑𝑗𝑎𝑐𝑒𝑛𝑡 𝑎𝑟𝑒𝑎𝑠 𝑖𝑛𝑠𝑖𝑑𝑒 𝑡ℎ𝑒 𝑠𝑎𝑚𝑒 ℎ𝑜𝑢𝑠𝑒
+ ∑ ∑ (𝐴𝑙𝑚𝑖ℎ 𝑇𝑚
ℎ(𝑘))
𝑁𝑑ℎ
𝑚=1
𝑁𝑠
ℎ=1(ℎ≠𝑖)⏟ 𝑡ℎ𝑒𝑟𝑚𝑎𝑙 𝑐𝑜𝑛𝑡𝑟𝑖𝑏𝑢𝑡𝑖𝑜𝑛𝑠 𝑓𝑟𝑜𝑚 𝑎𝑑𝑗𝑎𝑐𝑒𝑛𝑡 𝑎𝑟𝑒𝑎𝑠 𝑓𝑟𝑜𝑚 𝑜𝑡ℎ𝑒𝑟 ℎ𝑜𝑢𝑠𝑒𝑠
+ 𝑣𝑙𝑖(𝑘),
(3.5)
where
𝐴𝑙𝑙𝑖𝑖 = (1 −
𝛥𝑡
𝑅𝑙𝑒𝑞𝑖 𝐶𝑙𝑒𝑞
𝑖) , 𝐵𝑙
𝑖 =𝛥𝑡
𝐶𝑙𝑒𝑞𝑖, 𝐷𝑙
𝑖 = ∑𝑇𝑔𝑖 − 𝑇𝑙
𝑖
𝑅𝑙𝑔𝑒𝑞𝑖 𝐶𝑙𝑒𝑞
𝑖𝛥𝑡
𝑁𝑑𝑖
𝑔=1(𝑔≠𝑙)
+ ∑ ∑𝑇𝑚ℎ − 𝑇𝑙
𝑖
𝑅𝑙𝑚𝑒𝑞𝑖 𝐶𝑙𝑒𝑞
𝑖𝛥𝑡
𝑁𝑑ℎ
𝑚=1
𝑁𝑠
ℎ=1(ℎ≠𝑖)
,
𝑣𝑙𝑖 =
𝑃𝑙𝑃𝑑𝑖 𝛥𝑡
𝐶𝑙𝑒𝑞𝑖
+𝑇𝑜𝑎𝛥𝑡
𝑅𝑙𝑒𝑞𝑖 𝐶𝑙𝑒𝑞
𝑖.
(3.6)
where Ns is number of TCA’s, Ndi is number of divisions inside subsystem (i), 𝑥𝑙𝑖 is the indoor
temperature in TCA/subsystem (i) inside division (l), 𝑢𝑙𝑖 is the used power to provide comfort in
TCA/subsystem (i) inside division (l), 𝐴𝑙𝑚𝑖ℎ is an element from the state matrix A that relates the
state (indoor temperature) in division (m) from TCA/subsystem (h), with the state from division
(l) in TCA (i) and 𝑣𝑙𝑖 is the thermal disturbance in TCA/subsystem (i) inside division (l) e(g.
load generated by occupants, direct sunlight, electrical devices or doors and windows aperture
to recycle the indoor air), and Toa, the temperature of outside air (ºC). Remark that the number
of states variables in general model (3.4) is
SELECTION AND IDENTIFICATION SCENARIOS AND MODELS
62
𝑛 =∑ 𝑁𝑑𝑖𝑁𝑠
𝑖=1.
To exemplify, Figure 3.10 illustrates a house/TCA with six divisions that thermally interact
between them.
Figure 3.10. TCA divisions interaction simplified scheme;
So, for each division, (3.4) can be written as:
𝑥11(𝑘 + 1) = 𝐴11
11𝑥11(𝑘) + 𝐴12
11𝑥21(𝑘) + 𝐵1
1𝑢11(𝑘) + 𝑣1
1(𝑘),
𝑥21(𝑘 + 1) = 𝐴22
11𝑥21(𝑘) + 𝐴21
11𝑥11(𝑘) + 𝐴23
11𝑥31(𝑘) + 𝐴24
11𝑥41(𝑘) + 𝐴26
11𝑥61(𝑘) + 𝐵2
1𝑢21(𝑘) + 𝑣2
1(𝑘),
𝑥31(𝑘 + 1) = 𝐴33
11𝑥31(𝑘) + 𝐴32
11𝑥21(𝑘) + 𝐴35
11𝑥51(𝑘) + 𝐴36
11𝑥61(𝑘) + 𝐵3
1𝑢31(𝑘) + 𝑣3
1(𝑘), (3.7)
𝑥41(𝑘 + 1) = 𝐴44
11𝑥41(𝑘) + 𝐴42
11𝑥21(𝑘) + 𝐴45
11𝑥51(𝑘) + 𝐴46
11𝑥61(𝑘) + 𝐵4
1𝑢41(𝑘) + 𝑣4
1(𝑘),
𝑥51(𝑘 + 1) = 𝐴55
11𝑥51(𝑘) + 𝐴53
11𝑥31(𝑘) + 𝐴54
11𝑥41(𝑘) + 𝐴56
11𝑥61(𝑘) + 𝐵5
1𝑢51(𝑘) + 𝑣5
1(𝑘),
𝑥61(𝑘 + 1) = 𝐴66
11𝑥61(𝑘) + 𝐴62
11𝑥21(𝑘) + 𝐴63
11𝑥31(𝑘) + 𝐴64
11𝑥41(𝑘) + 𝐴65
11𝑥51(𝑘) + 𝐵6
1𝑢61(𝑘) + 𝑣6
1(𝑘),
dl1 dl2
dl4 dl3
dl5
dl6
SELECTION AND IDENTIFICATION SCENARIOS AND MODELS
63
[ 𝑥11(𝑘 + 1)
𝑥21(𝑘 + 1)
𝑥31(𝑘 + 1)
𝑥41(𝑘 + 1)
𝑥51(𝑘 + 1)
𝑥61(𝑘 + 1)]
=
[ 𝐴1111 0 0 0 0 0
0 𝐴2211 0 0 0 0
0 0 𝐴3311 0 0 0
0 0 0 𝐴4411 0 0
0 0 0 0 𝐴5511 0
0 0 0 0 0 𝐴6611]
[ 𝑥11(𝑘)
𝑥21(𝑘)
𝑥31(𝑘)
𝑥41(𝑘)
𝑥51(𝑘)
𝑥61(𝑘)]
+
[ 0 𝐴12
11 0 0 0 0
𝐴2111 0 𝐴23
11 𝐴2411 0 𝐴26
11
0 𝐴3211 0 0 𝐴35
11 𝐴3611
0 𝐴4211 0 0 𝐴45
11 𝐴4611
0 0 𝐴5311 𝐴54
11 0 𝐴5611
0 𝐴6211 𝐴63
11 𝐴6411 𝐴65
11 0 ]
[ 𝑥11(𝑘)
𝑥21(𝑘)
𝑥31(𝑘)
𝑥41(𝑘)
𝑥51(𝑘)
𝑥61(𝑘)]
+
[ 𝐵11 0 0 0 0 0
0 𝐵21 0 0 0 0
0 0 𝐵31 0 0 0
0 0 0 𝐵41 0 0
0 0 0 0 𝐵51 0
0 0 0 0 0 𝐵61]
[ 𝑢11(𝑘)
𝑢21(𝑘)
𝑢31(𝑘)
𝑢41(𝑘)
𝑢51(𝑘)
𝑢61(𝑘)]
+
[ 𝑣11(𝑘)
𝑣21(𝑘)
𝑣31(𝑘)
𝑣41(𝑘)
𝑣51(𝑘)
𝑣61(𝑘)]
.
(3.8)
For a more complex scenario, Figure 3.11 shows a distributed environment with five houses
with different plans and with different zones that may thermally interact. Remark that as
described in the sequel, adjacent areas can be doubly coupled, thermally and by the power
constraint. In Figure 3.11, ui represents the input (heat/cooling power) and yi is the output vector
containing the temperatures/states inside the several divisions of house (i)
Figure 3.11. Generalized house/TCA scheme example.
Thus, (3.4) can be generalized and presented as:
𝑥11(𝑘 + 1) = 𝐴11
11𝑥11(𝑘) + 𝐴12
11𝑥21(𝑘) + 𝐴13
11𝑥31(𝑘) + 𝐴11
12𝑥12(𝑘)+𝐴12
12𝑥22(𝑘)+𝐵1
1𝑢11(𝑘) + 𝑣1
1(𝑘),
𝑥21(𝑘 + 1) = 𝐴22
11𝑥21(𝑘) + 𝐴21
11𝑥11(𝑘) + 𝐴23
11𝑥31(𝑘)+𝐵2
1𝑢21(𝑘) + 𝑣(𝑘),
𝑥31(𝑘 + 1) = 𝐴33
11𝑥31(𝑘) + 𝐴32
11𝑥21(𝑘) + 𝐴31
11𝑥11(𝑘) + 𝐴32
12𝑥22(𝑘)+𝐵3
1𝑢31(𝑘) + 𝑣3
1(𝑘),
𝑥12(𝑘 + 1) = 𝐴11
22𝑥12(𝑘) + 𝐴12
22𝑥22(𝑘) + 𝐴11
21𝑥11(𝑘)+𝐵1
2𝑢12(𝑘) + 𝑣1
2(𝑘),
𝑥22(𝑘 + 1) = 𝐴22
22𝑥22(𝑘) + 𝐴21
22𝑥12(𝑘) + 𝐴21
21𝑥11(𝑘) +
𝐴2321𝑥3
1(𝑘)+𝐴2223𝑥2
3(𝑘) + 𝐴2124𝑥1
4(𝑘) + 𝐵22𝑢2
2(𝑘) + 𝑣22(𝑘),
(3.9)
1
2
43
1u
1y
2u
2y
3u
3y4u
4y
55u
5yd11
d12
d13
d22d31
d32
d41
d21w3
w2
w1
w4d51
w5
SELECTION AND IDENTIFICATION SCENARIOS AND MODELS
64
𝑥13(𝑘 + 1) = 𝐴11
33𝑥13(𝑘) + 𝐴12
33𝑥23(𝑘) + 𝐴11
34𝑥14(𝑘)+𝐵1
3𝑢13(𝑘) + 𝑣1
3(𝑘),
𝑥23(𝑘 + 1) = 𝐴22
33𝑥23(𝑘) + 𝐴21
33𝑥13(𝑘) + 𝐴21
34𝑥14(𝑘)𝐴21
32𝑥12(𝑘)+𝐴22
32𝑥22(𝑘) + 𝐵2
3𝑢23(𝑘) + 𝑣2
3(𝑘),
𝑥14(𝑘 + 1) = 𝐴11
44𝑥14(𝑘) + 𝐴11
43𝑥13(𝑘) + 𝐴12
43𝑥23(𝑘)𝐴12
42𝑥22(𝑘) + 𝐵1
4𝑢14(𝑘) + 𝑣1
4(𝑘),
𝑥15(𝑘 + 1) = 𝐴11
55𝑥15(𝑘)+𝐵1
5𝑢15(𝑘) + 𝑣1
5(𝑘),
[ 𝑥11(𝑘 + 1)
𝑥21(𝑘 + 1)
𝑥31(𝑘 + 1)
𝑥12(𝑘 + 1)
𝑥22(𝑘 + 1)
𝑥13(𝑘 + 1)
𝑥23(𝑘 + 1)
𝑥14(𝑘 + 1)
𝑥15(𝑘 + 1)]
=
[ 𝐴1111 𝐴12
11 𝐴1311 𝐴11
12 𝐴1212 0 0 0 0
𝐴2111 𝐴22
11 𝐴2311 0 0 0 0 0 0
𝐴3111 𝐴31
11 𝐴3311 𝐴32
12 0 0 0 0 0
𝐴1121 0 0 𝐴11
22 𝐴1222 0 0 0 0
𝐴2121 0 𝐴23
21 𝐴2121 𝐴22
22 0 𝐴2223 𝐴21
24 0
0 0 0 0 0 𝐴1133 𝐴12
33 𝐴1134 0
0 0 0 𝐴2132 𝐴22
32 𝐴2133 𝐴22
33 𝐴2134 0
0 0 0 𝐴1242 0 𝐴11
43 𝐴1243 𝐴11
44 0
0 0 0 0 0 0 0 0 𝐴1155]
[ 𝑥11(𝑘)
𝑥21(𝑘)
𝑥31(𝑘)
𝑥12(𝑘)
𝑥22(𝑘)
𝑥13(𝑘)
𝑥23(𝑘)
𝑥14(𝑘)
𝑥15(𝑘)]
+
+
[ 𝐵11 0 0 0 0 0 0 0 0
0 𝐵21 0 0 0 0 0 0 0
0 0 𝐵31 0 0 0 0 0 0
0 0 0 𝐵12 0 0 0 0 0
0 0 0 0 𝐵22 0 0 0 0
0 0 0 0 0 𝐵13 0 0 0
0 0 0 0 0 0 𝐵23 0 0
0 0 0 0 0 0 0 𝐵14 0
0 0 0 0 0 0 0 0 𝐵15]
[ 𝑢11(𝑘)
𝑢21(𝑘)
𝑢31(𝑘)
𝑢12(𝑘)
𝑢22(𝑘)
𝑢13(𝑘)
𝑢23(𝑘)
𝑢14(𝑘)
𝑢15(𝑘)]
+
[ 𝑣11(𝑘)
𝑣21(𝑘)
𝑣31(𝑘)
𝑣12(𝑘)
𝑣22(𝑘)
𝑣13(𝑘)
𝑣23(𝑘)
𝑣14(𝑘)
𝑣15(𝑘)]
.
65
Chapter 4
4 MPC and PI control in thermal
comfort systems
4.1 Introduction
The chapter presents the first approach towards the final developed control structure. It allowed
us to understand the problem, its dynamics and to verify that the MPC control strategy is
suitable for the objective that is intent to be achieved. So, it considers one house/TCA and no
communication with neighbouring is taking into account (Barata et al., 2012a).
The methodologies are based on a predictive control structure associated to a PI controller and
applied to the thermal house comfort in an environment with limited energy sources. Thus, the
available and necessary power to cool and heat the house within the desire “comfort zone” is
partially available and, consequently constrained. The system also uses the house thermal inertia
to “accommodate thermal energy” within the selected temperature range, maintaining the
comfort when no energy is available and reducing the energy consumption without loss of the
thermal comfort. The simulation results obtained with traditional PI and with MPC are
presented and it can be seen that the system is able to provide high-energy conservation and
reliable comfort. The house temperature variations are calculated and are taking into account the
necessary heat and cooling needs and heat losses to the environment. The indoor temperature
time derivative is expressed as follows,
𝑑𝑇𝑙𝑖
𝑑𝑡=
1
𝐶𝑙𝑒𝑞𝑖(𝑄𝑙
𝑖ℎ𝑒𝑎𝑡
− 𝑄𝑙𝑙𝑜𝑠𝑠𝑒𝑠𝑖 + 𝑄𝑙𝑃𝑑
𝑖 ). (4.1)
MPC AND PI CONTROL IN THERMAL COMFORT SYSTEMS
66
Actually (4.1) is not an accurate model of the house. Nevertheless, it reflects well enough the
dominant model of a real house to be used for designing the MPC controller.
The heat loss is mainly a function of outdoor air temperature. By taking the outdoor temperature
as a primary influencing factor for the weather, the heat loss is given by,
𝑄𝑙𝑙𝑜𝑠𝑠𝑒𝑠𝑖 =
𝑇𝑜𝑢𝑡 − 𝑇𝑙𝑖
𝑅𝑙𝑒𝑞𝑖
. (4.2)
The parameter 𝑅𝑙𝑒𝑞𝑖 describes the equivalent thermal resistance of all walls (including roof and
ceiling) and windows that isolate the house from outside, and can be describe as a electrical
parallel resistance circuit (Gulley and Chistol, 2006; Ma, et al., 2012).
𝑅𝑙𝑒𝑞𝑖 = 𝑅𝑙𝑤𝑎𝑙𝑙𝑠
𝑖 //𝑅𝑙𝑤𝑖𝑛𝑑𝑜𝑤𝑠𝑖 . (4.3)
As also used in (Löhnberg, 1999; Mendes, et al., 2001), the external temperature disturbance
variation, Tout, can be approximated represented by a sine (with a given amplitude) where is
added a constant temperature value. Other disturbances can be reflected in (4.1), like heat flux
from solar radiation, inside loads generated by occupancy, lights or other electrical devices (Ma,
et al., 2012).
The plant model representation (4.1) can be rewritten and changed into a discrete model.
𝑦(𝑘 + 1) = 𝑎𝑦(𝑘) + 𝑏𝑢(𝑘) + 𝑐𝑣(𝑘), (4.4)
where, 𝑎 = (1 −Δt
𝑅𝑙𝑒𝑞𝑖 𝐶𝑙𝑒𝑞
𝑖 ) , 𝑏 =Δt
𝐶𝑙𝑒𝑞𝑖 , c=
Δt
𝑅𝑙𝑒𝑞𝑖 𝐶𝑙𝑒𝑞
𝑖 , u(k) is the necessary heat/cooling power, v(k)
are the disturbances and y(k) is the indoor temperature. Note that the space is cooled when u(k)
< 0 and heated when u(k) > 0.
4.2 System description
As a first approach towards developing a control structures it is considered an individual house,
Figure 4.1. As mentioned, this approach does not take into account the possibility of
communication with neighbouring houses.
MPC AND PI CONTROL IN THERMAL COMFORT SYSTEMS
67
PI
+
HVACMPC
-+
r0(k) u(k) y(k)
v(k)
Disturbances forecasts
Environment temperature,
solar loadsEnvironment temperature,
solar loads
Disturbances
Avaiable power
Figure 4.1. Block diagram of the implemented system.
The following discrete equations describe the implemented system showed in Figure 4.1.
𝑌(𝑧) =𝑏
𝑧 − 𝑎𝑈(𝑧) +
𝑐
𝑧 − 𝑎𝑉(𝑧). (4.5)
With the controller described by,
𝑃𝐼 = 𝑘𝑝 +𝑘𝑖𝑧
𝑧 − 1=𝑧(𝑘𝑝 + 𝑘𝑖) − 𝑘𝑝
𝑧 − 1. (4.6)
Being
𝑈(𝑧) = 𝑘𝑧 − 𝛽
𝑧 − 1𝐸(𝑧), (4.7)
and
𝐸(𝑧) = 𝑅0(𝑧) − 𝑌(𝑧). (4.8)
Figure 4.1 can also be characterized by the following discrete equations, where (4.9) and (4.10)
describe the system and (4.11) and (4.12) the controller.
𝑥(𝑘 + 1) = 𝐴𝑥(𝑘) + 𝐵𝑢(𝑘) + 𝐸𝑣(𝑘), (4.9)
𝑦(𝑘) = 𝐶𝑥(𝑘), (4.10)
𝜉(𝑘 + 1) = 𝐴𝑐𝜉(𝑘) + 𝐵𝑐[𝑟(𝑘) − 𝑦(𝑘)], (4.11)
𝑢(𝑘) = 𝐶𝑐𝜉(𝑘). (4.12)
Substituting (4.12) in (4.9) and (4.10) in (4.11) results in (4.13) and (4.14), respectively that
describe the close-loop system in u(k).
𝑥(𝑘 + 1) = 𝐴𝑥(𝑘) + 𝐵𝐶𝑐𝜉(𝑘) + 𝐸𝑣(𝑘), (4.13)
MPC AND PI CONTROL IN THERMAL COMFORT SYSTEMS
68
𝜉(𝑘 + 1) = 𝐴𝑐𝜉(𝑘) + 𝐵𝑐[𝑟(𝑘) − 𝐶𝑥(𝑘)], (4.14)
[𝑥(𝑘 + 1)
𝜉(𝑘 + 1)] = [
𝐴 𝐵𝐶𝑐−𝐵𝑐𝐶 𝐴𝑐
] [𝑥(𝑘)
𝜉(𝑘)] + [
0𝐵𝑐] 𝑟(𝑘) + [
𝐸0] 𝑣(𝑘). (4.15)
The following equations describe the close-loop system ∆r(k)
𝑟(𝑘) = 𝛥𝑟(𝑘) + 𝑟(𝑘 − 1), (4.16)
𝑥(𝑘 + 1) = 𝐴𝑥(𝑘) + 𝐵𝐶𝑐𝜉(𝑘) + 𝐸𝑣(𝑘). (4.17)
Equation (4.18) results from substituting (4.16) in (4.14).
𝜉(𝑘 + 1) = 𝐴𝑐𝜉(𝑘) − 𝐵𝑐𝐶𝑥(𝑘) + 𝐵𝑐[𝛥𝑟(𝑘) − 𝑟(𝑘 − 1)], (4.18)
with
𝛾(𝑘) = 𝑟(𝑘 − 1), (4.19)
𝛾(𝑘 + 1) = 𝛾(𝑘) + 𝛥𝑟(𝑘), (4.20)
𝑥(𝑘 + 1) = 𝐴𝑥(𝑘) + 𝐵𝐶𝑐𝜉(𝑘) + 𝐸𝑣(𝑘). (4.21)
Substituting (4.20) in (4.18) results in
𝜉(𝑘 + 1) = −𝐵𝑐𝐶𝑥(𝑘) + 𝐴𝑐𝜉(𝑘) + 𝐵𝑐𝛾(𝑘) + 𝐵𝑐𝛥𝑟(𝑘), (4.22)
[
𝑥(𝑘 + 1)
𝜉(𝑘 + 1)
𝛾(𝑘 + 1)] = [
𝐴 𝐵𝐶𝑐 0−𝐵𝑐𝐶 𝐴𝑐 𝐵𝐶 0 0 𝐼
]
⏞ 𝐴1
[
𝑥(𝑘)𝜉(𝑘)
𝛾(𝑘)]
⏞ 𝜁(𝑘)
+ [0 𝐸𝐵𝐶 0𝐼 0
]
⏞ [𝐵1 𝐵2]
[𝛥𝑟0(𝑘)𝑣(𝑘)
], (4.23)
𝑦(𝑘) = [𝐶 0 0⏟ 𝐶𝑦
] [
𝑥(𝑘)
𝜉(𝑘)
𝛾(𝑘)], (4.24)
𝑢(𝑘) = [0 𝐶𝑐 0⏟ 𝐶𝑢
] [
𝑥(𝑘)𝜉(𝑘)
𝛾(𝑘)], (4.25)
𝑟 = [0 0 𝐼] [
𝑥(𝑘)
𝜉(𝑘)
𝛾(𝑘)], (4.26)
Or,
𝜉(𝑘 + 1) = 𝐴1𝜁(𝑘) + 𝐵1𝛥𝑟0(𝑘) + 𝐵2𝑣(𝑘), (4.27)
MPC AND PI CONTROL IN THERMAL COMFORT SYSTEMS
69
𝑦(𝑘) = 𝐶𝑦𝜁(𝑘) 𝑢(𝑘) = 𝐶𝑢𝜁(𝑘). (4.28)
As mentioned in Chapter 2.3 predictive control is based on the receding horizon principle in
which, at each sample interval, the current controller output is obtained by solving an optimal
control problem of finite horizon. The input to the controller is current output/state information
and the output is a sequence of future control actions where usually the first element in the
sequence is applied to the plant.
Predictive control belongs to the class of model-based designs where a model of the plant is
used to predict the behaviour of the plant and calculate the controller output such that a given
control criterion is minimized.
The criterion must be selected such that the performance and robustness specifications are
accomplished. The GPC criterion introduced in (Clarke et al., 1987) is of the form:
𝐽 = ∑[∑‖𝑦𝑖(𝑘 + 𝑙) − 𝑦𝑖𝑑(𝑘 + 𝑙)‖
𝑄𝑖
2+
𝐻𝑃
𝑙=1
∑‖𝛥𝑢𝑖(𝑘 + 𝑙 − 1)‖𝑅𝑖2
𝑁𝐶
𝑙=1
]
𝑛
𝑖=1
, (4.29)
where Qi and Ri are weight matrices, HP, NC are predictive horizon and control horizon,
respectively, and HP ≥ M, yid is the set-point of subsystem Si , Δui(k) = ui(k) − Δui(k − 1) is
the input increment vector of subsystem Si .
Considering Figure 4.1, (4.29) can be rewritten,
𝐽 =∑[∑‖𝑦𝑖(𝑘 + 𝑙) − 𝑟𝑖𝑑(𝑘 + 𝑙)‖
𝑄𝑖
2+
𝐻𝑃
𝑙=1
∑‖𝛥𝑟𝑜𝑖(𝑘 + 𝑙 − 1)‖𝑅𝑖2
𝑁𝐶
𝑙=1
]
𝑛
𝑖=1
. (4.30)
Then the model outputs in future sample intervals are given as follows:
𝑦 = [
𝑦 (𝑘 + 1)
𝑦 (𝑘 + 2)⋮
𝑦 (𝑘 + 𝐻𝑃)
] = [
𝑦 𝑜(𝑘 + 1)
𝑦 𝑜(𝑘 + 2)⋮
𝑦 𝑜(𝑘 + 𝐻𝑃)
] + [
𝑔1 0 … 0𝑔2 𝑔1 ⋱ ⋮⋮ ⋮ ⋱ 0𝑔𝐻𝑃 𝑔𝐻𝑃−1 … 𝑔𝐻𝑃−𝑁𝐶−1
]
× [
∆𝑟𝑜(𝑘)
∆𝑟𝑜(𝑘 + 1)⋮
∆𝑟𝑜(𝑘 + 𝑁𝐶 − 1)
].
(4.31)
Where,
MPC AND PI CONTROL IN THERMAL COMFORT SYSTEMS
70
[
𝑦 𝑜(𝑘 + 1)
𝑦 𝑜(𝑘 + 2)⋮
𝑦 𝑜(𝑘 + 𝐻𝑃)
] =
[ 𝐶1𝐴1 𝐶2𝐴2 … 𝐶𝑛𝐴𝑛𝐶1𝐴1
2 𝐶2𝐴22 ⋱ 𝐶𝑛𝐴𝑛
2
⋮ ⋮ ⋱ ⋮𝐶1𝐴1
𝐻𝑃 𝐶2𝐴2𝐻𝑃 … 𝐶𝑛𝐴𝑛
𝐻𝑃] [
𝑥𝑜(𝑘)
𝑥𝑜(𝑘 + 1)⋮
𝑥𝑜(𝑘 + 𝑁𝐶 − 1)
] +
[
𝑤1 0 … 0𝑤1 𝑤1 ⋱ ⋮⋮ ⋮ ⋱ 0𝑤𝐻𝑃 𝑤𝐻𝑃−1 … 𝑤𝐻𝑃−𝑁𝐶−1
] [
𝑣(𝑘)
𝑣(𝑘 + 1)⋮
𝑣(𝑘 + 𝑁𝐶 − 1)
].
(4.32)
In compact form, (4.31) and (4.32) can be written as the following relations:
�� = ��𝑜 + 𝐺∆𝑅𝑜, (4.33)
��0 = 𝐶𝐴𝑥0 +𝑊𝑣. (4.34)
Being the prediction error e determined using the following equation,
𝑒 = 𝑦 − 𝑟𝑑 , (4.35)
where rd is the set-point. Substituting (4.33) in (4.35), and the result in (4.30), the objective
function can be presented in the form,
𝐽(∆𝑅𝑜) = (𝑌�� + 𝐺∆𝑅𝑜 − 𝑅𝑑)𝑇𝑄(��𝑜 + 𝐺∆𝑅𝑜 − 𝑅𝑑) + ∆𝑅𝑜
𝑇𝑅∆𝑅𝑜. (4.36)
The solution minimising J to give the optimal suggested control increment 𝑅𝑜∗, in the absence of
constrains, is given by 𝜕𝐽
𝜕Δ𝑅𝑜, resulting,
∆𝑅𝑜∗ = (𝐺𝑇𝑄𝐺 + 𝑅)−1𝐺𝑇𝑄(𝑅𝑑 − 𝑌��). (4.37)
To simulate the system it was considered, that the house dynamics is represented by (4.38) and
the PI controller by (4.39), respectively.
𝐾𝑃𝑙𝑎𝑛𝑡𝜏𝑃𝑙𝑎𝑛𝑡 𝑠 + 1
, (4.38)
𝐾𝑃𝐼 (1 +1
𝜏𝑃𝐼 𝑠). (4.39)
With 𝐾𝑃𝑙𝑎𝑛𝑡 and τ𝑃𝑙𝑎𝑛𝑡 are the house and 𝐾𝑃𝐼 and 𝜏𝑃𝐼the controller parameters. Applying the
above FT’s in Figure 4.1, and analysing in separated each one of the entries, with and without
perturbation, results the in next equations (4.40) and (4.41), respectivly.
𝐾𝑃𝑙𝑎𝑛𝑡𝜏𝑃𝐼 𝑠
𝜏𝑃𝐼 𝜏𝑃𝑙𝑎𝑛𝑡 𝑠2 + 𝜏𝑃𝐼 𝑠(1 + 𝐾𝑃𝑙𝑎𝑛𝑡𝐾𝑃𝐼) + 𝐾𝑃𝑙𝑎𝑛𝑡𝐾𝑃𝐼
, (4.40)
MPC AND PI CONTROL IN THERMAL COMFORT SYSTEMS
71
𝐾𝑃𝑙𝑎𝑛𝑡𝐾𝑃𝐼(1 + 𝜏𝑃𝐼 𝑠)
𝜏𝑃𝐼 𝜏𝑃𝑙𝑎𝑛𝑡 𝑠2 + 𝜏𝑃𝐼 𝑠(1 + 𝐾𝑃𝑙𝑎𝑛𝑡𝐾𝑃𝐼) + 𝐾𝑃𝑙𝑎𝑛𝑡𝐾𝑃𝐼
. (4.41)
Converting the system to discrete space–state model, the augmented space-state model is given
by Ai , Bi and Ci. resulting matrixes are presented in the following format.
[𝑥(𝑘 + 1)𝑢(𝑘)
] = [𝐴 𝐵0 𝐼
] [𝑥(𝑘)
𝑢(𝑘 − 1)] + [
𝐵𝐼] ∆𝑢(𝑘), (4.42)
𝑦(𝑘) = [𝐶 0] [𝑥(𝑘)
𝑢(𝑘 − 1)], (4.43)
where A, B and C results from continuous to discrete space-state transformation using Zero
Order Hold (ZOH) (Rugh, 1996).
According to the above showed, the equations (4.33) and (4.34) parameters can assume now the
final form and can be written as follow.
𝐶𝐴 =
[ 𝐶𝑖𝐴𝑖 𝐶𝑖𝐴𝑖 … 𝐶𝑖𝐴𝑖𝐶𝑖𝐴𝑖
2 𝐶𝑖2𝐴𝑖2 ⋱ 𝐶𝑖𝐴𝑖
2
⋮ ⋮ ⋱ ⋮𝐶𝑖𝐴𝑖
𝐻𝑃 𝐶𝑖𝐴𝑖𝐻𝑃 … 𝐶𝑖𝐴𝑖
𝐻𝑃] , (4.44)
𝐺 =
[
𝐶𝑖1𝐵1𝑖 0 0 … … 0𝐶𝑖𝐴𝑖𝐵1𝑖 𝐶𝑖1𝐵1𝑖 0 … … 0
𝐶𝑖𝐴𝑖2𝐵1𝑖 𝐶𝑖𝐴𝑖𝐵1𝑖 𝐶𝑖1𝐵1𝑖 0 … ⋮⋮ ⋮ ⋮ … … ⋮
𝐶𝑖𝐴𝑖𝐻𝑃−2𝐵1𝑖 𝐶𝑖𝐴𝑖
𝐻𝑃−3𝐵1𝑖 𝐶𝑖𝐴𝑖𝐻𝑃−4𝐵1𝑖 … 0 0
𝐶𝑖𝐴𝑖𝐻𝑃−1𝐵1𝑖 𝐶𝑖𝐴𝑖
𝐻𝑃−2𝐵1𝑖 𝐶𝑖𝐴𝑖𝐻𝑃−3𝐵1𝑖 … 𝐶𝑖𝐴𝑖𝐵1𝑖 𝐶𝑖1𝐵1𝑖]
, (4.45)
𝑊 =
[
𝐶𝑖1𝐵2𝑖 0 0 … … 0𝐶𝑖𝐴𝑖𝐵2𝑖 𝐶𝑖1𝐵2𝑖 0 … … 0
𝐶𝑖𝐴𝑖2𝐵2𝑖 𝐶𝑖𝐴𝑖𝐵2𝑖 𝐶𝑖1𝐵2𝑖 0 … ⋮⋮ ⋮ ⋮ … … ⋮
𝐶𝑖𝐴𝑖𝐻𝑃−2𝐵2𝑖 𝐶𝑖𝐴𝑖
𝐻𝑃−3𝐵2𝑖 𝐶𝑖𝐴𝑖𝐻𝑃−4𝐵2𝑖 … 0 0
𝐶𝑖𝐴𝑖𝐻𝑃−1𝐵2𝑖 𝐶𝑖𝐴𝑖
𝐻𝑃−2𝐵2𝑖 𝐶𝑖𝐴𝑖𝐻𝑃−3𝐵2𝑖 … 𝐶𝑖𝐴𝑖𝐵2𝑖 𝐶𝑖1𝐵2𝑖]
, (4.46)
where B1i and B2i emerge from matrix Bi corresponding each one to one of the system inputs.
Rewriting (4.33) and (4.34) results in,
�� = 𝑌�� + 𝐺∆𝑅0 , (4.47)
𝑌�� = 𝛤𝑦𝜁(𝑘) +𝑊𝑦𝑣. (4.48)
From matrices T(C, ��) and L(C, ��, ��) below,
MPC AND PI CONTROL IN THERMAL COMFORT SYSTEMS
72
𝑻(��, ��) = [
��������2
⋮����𝐻𝑃
], (4.49)
ℒ(��, ��, ��)=
[
���� 0 … 0������ ���� … 0����2�� ������ 0 ⋮⋮ ⋮ … ⋮
����𝐻𝑃−2�� ����𝐻𝑃−3�� … 0����𝐻𝑃−1�� ����𝐻𝑃−2�� … ����]
. (4.50)
The matrices G, Wy and Γy can be obtained and formalized as:
𝛤𝑦 = T(𝐶𝑦, 𝐴1), 𝛤𝑢 = T(𝐶𝑢 , 𝐴1), (4.51)
𝑊𝑦 = ℒ(𝐶𝑦 , 𝐴1, 𝐵2) and 𝑊𝑢 = ℒ(𝐶𝑢, 𝐴1, 𝐵2), (4.52)
𝐺 = ℒ(𝐶𝑦 , 𝐴1, 𝐵1), 𝐺𝑢 = ℒ(𝐶𝑢 , 𝐴1, 𝐵1). (4.53)
As mentioned above, the control is now made with constrains to find the r0(k) that minimize
the cost function presented in (4.54). Assuming that the available energy is conditioned, the
control method intent to restrict the value of the heating/cooling user power demanding, u. The
designed, Figure 4.2 system also includes constrains in upper and lower value of r0 and r0.
1B(q-1)1+ + y(k)
v(k)
A(q-1)AC(q-1)
B2C(q-1)
B1C(q-1)RG
r(k) r0(k)
+
-
u(k)
+
HomeHome
Controller
Governor 1
Figure 4.2. MPC with constraints, system block diagram.
The predictive controller can be posed as a quadratic programming problem, where,
𝑚𝑖𝑛∆𝑅0
𝐽(𝑘) = (𝑌�� + 𝐺∆𝑅0 − 𝑅𝑑)𝑇𝑄𝑒(𝑌�� + 𝐺∆𝑅0 − 𝑅𝑑) + ∆𝑅0
𝑇𝑅𝑒∆𝑅0, (4.54)
MPC AND PI CONTROL IN THERMAL COMFORT SYSTEMS
73
is to be optimized, subject to the constraints (4.55)-(4.59):
𝑦(𝑘) = 𝐴(𝑞−1)𝑦(𝑘 − 1) + 𝐵(𝑞−1)𝑢(𝑘 − 1) + 𝐶(𝑞−1)𝑣(𝑘 − 1), (4.55)
𝑢(𝑘) = 𝐴𝑐(𝑞−1)𝑢(𝑘 − 1) + 𝐵1𝑐(𝑞
−1)𝑟0(𝑘) − 𝐵2𝑐(𝑞−1)𝑦(𝑘), (4.56)
𝑟𝑜(𝑘) ≤ 𝑟𝑜(𝑘) ≤ 𝑟𝑜(𝑘), (4.57)
∆𝑟𝑜(𝑘) ≤ ∆𝑟𝑜(𝑘) ≤ ∆𝑟𝑜(𝑘), (4.58)
𝑢(𝑘) ≤ 𝑢(𝑘) ≤ 𝑢(𝑘). (4.59)
Developing (4.54), the new cost function can be written as (Igreja and Cruces, 2002)
𝑚𝑖𝑛∆𝑅0
𝐽 = (1
2∆𝑅0
𝑇𝑄𝑢∆𝑅0) + 𝐶𝑢𝑇∆𝑅0, (4.60)
with
𝑄𝑢 = 2(𝐺𝑇𝑄𝑒𝐺 + 𝑅𝑒), (4.61)
And
𝐶𝑢𝑇 = −2(𝑅𝑑 − 𝑌��)
𝑇𝑄𝑒 . 𝐺. (4.62)
The constrains can be expressed in terms of control movements resulting,
𝑅0 ≤ 𝑇𝑙∆𝑅0 + [𝐼𝑚 … 𝐼𝑚]𝑟𝑜(𝑡 − 1) ≤ 𝑅0 , (4.63)
∆𝑅0 ≤ ∆𝑅 ≤ ∆𝑅0, (4.64)
𝑈 ≤ 𝐺∆𝑅0 + 𝑈�� ≤ 𝑈, (4.65)
where Tl is a lower triangular matrix whose non null submatrices are Im the identity matrix.
Introducing the new variable,
𝑋 = 𝑇𝑙∆𝑅𝑜 + 𝛱 . (4.66)
𝛱 = [𝐼𝑚 …𝐼𝑚]𝑟𝑜(𝑘 − 1) − 𝑅0. (4.67)
∆Ro becomes,
∆𝑅𝑜 = 𝑉(𝑋 − 𝛱), (4.68)
MPC AND PI CONTROL IN THERMAL COMFORT SYSTEMS
74
𝑉 = 𝑇𝑙−1. (4.69)
And consequently (4.60) becomes,
𝑚𝑖𝑛∆𝑅0
𝐽 = (1
2𝑋𝑇𝑄𝑋𝑋) + 𝐶𝑋
𝑇𝑋, (4.70)
𝑄𝑥 = 𝑉𝑇𝑄𝑢𝑉 = 2𝑉
𝑇(𝐺𝑇𝑄𝑒𝐺 + 𝑅𝑒)𝑉, (4.71)
and
𝐶𝑥𝑇 = 𝐶𝑢
𝑇 − 𝛱𝑇𝑉𝑇𝑄𝑢𝑉 = −2𝛱𝑇𝑉𝑇(𝐺𝑇𝑄𝑒𝐺 + 𝑅𝑒)𝑉. (4.72)
Finally, the constrains can be expressed as,
[ 𝐼−𝑉𝑉
−𝐺𝑢𝑉𝐺𝑢𝑉 ]
𝑋 ≤
[ 𝑅0 − 𝑅0
−∆𝑅0 − 𝑉𝛱
∆𝑅0 + 𝑉𝛱
−𝑈 − 𝐺𝑢𝑉𝛱 + ��𝑂
𝑈 + 𝐺𝑢𝑉𝛱 − ��𝑂 ]
. (4.73)
The QP solution can be obtained using a modified Lemke’s method (Camacho, 1993; Igreja and
Cruces, 2002).
Remark: the predicted needed power is given by (4.74).
��𝑂 = 𝛤𝑢𝜁(𝑘) +𝑊𝑢𝑣, (4.74)
where, the only active state in 𝛤𝑢 is the one that corresponds to u in ζ(𝑘). Matrices Wu and Γu
are showed in annex. With this formalization, constrains can be directly applied taking into
account the limited power variation along time.
4.3 Results
As mentioned, it was created a temperature “comfort zone”. This approach allows to weight
only the temperatures outside the gap, instead of traditional system were all deviations are
weight, and consequently, more resources are needed. Also, in order to reduce the energy needs,
it was defined that u(k)=0 inside the temperature bound. It is considered a comfort zone if the
indoor reference is eT relatively to the reference temperature. Two situations are here presented
and the obtained results can be seen in Chapter 4.3.2. In the first situation is considered that
outside the comfort zone, the system calculates u(k) and the cost function J, and, inside the
comfort zone both u and J are null. The second situation it is distinguished from the first one
MPC AND PI CONTROL IN THERMAL COMFORT SYSTEMS
75
because it considers that inside the comfort zone, the cost function J can always assume the
optimal value.
The generalised cost function can be written as (4.75).
𝐽(𝑘) =∑‖𝑒(𝑘 + 𝑙)‖𝑄𝑒2 +
𝐻𝑃
𝑙=1
∑‖𝛥𝑟0(𝑘 + 𝑙 − 1)‖𝑅𝑒2
𝑁𝐶
𝑙=1
], (4.75)
where Qe ≥0 and Re >0 are weight matrices and depend from eT., HP, NC are predictive
horizon and control horizon, respectively, with HP ≥ M and, Δr0(k) is the input increment
vector.
As mentioned, Qe and Re depend on eT value.
𝑄𝑒 {0 ‖𝑒‖ ≤ 𝑒𝑇𝑄 ‖𝑒‖ > 𝑒𝑇
(4.76)
𝑅𝑒 {0 ‖𝑒‖ ≤ 𝑒𝑇𝑅 ‖𝑒‖ > 𝑒𝑇
. (4.77)
In the simulated system it is considered that the house is exposed to an additional perturbation,
solar incidence, which elevates the outside temperature in more 7 ºC between 15 and 19 hours,
Figure 4.3.
Figure 4.3. Outdoor temperature.
The simulation period is 48 hours and the several approaches are compared. The results are
showed in the next subsections and the used parameters are in the table below.
5 10 15 20 25 30 35 40 4510
15
20
25
30
35
Time (h)
Tem
pera
ture
(ºC
)
Outdoor Temp.
MPC AND PI CONTROL IN THERMAL COMFORT SYSTEMS
76
Table 4.1.Simulation parameters for PIvsMPC
Parameter Value Units
HP 24 -
NC 48 -
∆t 0.5 h
KPlant 1 -
Plant 130 -
R 0.1 -
Q 1 -
PI gain (KPI) 200 -
PI time const (PI) 130 s
eT 0.5 ºC
4.3.1 PI versus MPC.
The results and performance with PI controller and with MPC+PI are here analysed, compared
and showed in Figure 4.4 to Figure 4.6.
Figure 4.4. Indoor temperature.
5 10 15 20 25 30 35 40 4518
18.5
19
19.5
20
20.5
21
21.5
22
Time (h)
Tem
pera
ture
(ºC
)
Temp. Ref.
Indoor Temp (PI)
Indoor Temp (MPC+PI)
Generated Ref (MPC+PI).
MPC AND PI CONTROL IN THERMAL COMFORT SYSTEMS
77
Figure 4.5. Power and energy consumption.
Figure 4.6. Temperature error with and without (MPC).
As can been seen in Figure 4.5 the total energy consumption is very similar, but, the MPC
provides the anticipative effect, Figure 4.4, that maintains the indoor temperature always in line
with the reference minimizing the error, Figure 4.6, comparatively with the PI solution.
4.3.2 MPC with “comfort zone”
As mentioned, here are present the obtained results of two distinguish situations. The “Case 1”
considered that outside the comfort zone u and the cost function are calculated and inside
comfort zone both u and J are null.
The second approach, “Case 2”, is similar to the first but the cost function J is always
calculated.
5 10 15 20 25 30 35 40 45-0.1
-0.05
0
0.05
0.1
0.15
0.2
0.25
0.3
Time (h)
Contr
ol In
put
(kW
) | E
nerg
y (
kW
h/1
0)
Control Input U (PI)
Control Input U (MPC+PI)
Energy(PI)
Energy (MPC+PI)
5 10 15 20 25 30 35 40 450
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
Time (h)
e2 (
ºC)
e2 (with MPC)
e2 (without MPC)
MPC AND PI CONTROL IN THERMAL COMFORT SYSTEMS
78
The next table summarize the actions that distinguish both approaches.
Table 4.2. Simplified algorithm for MPC with “comfort zone”.
Case 1 Case 2
if ‖𝑒‖ > 𝑒𝑇(0.5º𝐶) then compute u and J
else
u=0 and J=0
end if
if ‖𝑒‖ > 𝑒𝑇(0.5º𝐶) then compute u and J
else
u=0
compute J
end if
Figure 4.7 and Figure 4.9 show the results
Figure 4.7. Case 1 - Indoor temperature with “comfort zone”.
Figure 4.8. Case 2 - Indoor temperature with “comfort zone”.
5 10 15 20 25 30 35 40 4518
19
20
21
22
23
24
25
26
27
28
Time (h)
Te
mp
era
ture
(ºC
)
Temp. Ref.
Indoor Temp (Case 1)
Generated Ref (Case 1)
5 10 15 20 25 30 35 40 4518
19
20
21
22
23
24
25
26
27
28
Time (h)
Te
mp
era
ture
(ºC
)
Temp. Ref.
Indoor Temp (Case 2)
Generated Ref (Case 2)
MPC AND PI CONTROL IN THERMAL COMFORT SYSTEMS
79
Figure 4.7 shows that the generated reference “Case 1” is constant (no optimal increment is
calculated) when the difference between temperatures is within the range, and, due this fact, the
MPC anticipative effect is lost.
Figure 4.9. Power and energy consumption with “comfort zone”.
On the other hand in Figure 4.7 the anticipative effect is maintain in “Case 2”, and
consequently, Figure 4.9 reveals better results with less power “peaks” and also with less energy
expended comparatively with “Case 1”.
4.3.3 MPC with “comfort zone” and constrained available power
The results presented in this item consider that the available power is limited and variable.
Figure 4.10. Indoor temperature evolution with limited power.
5 10 15 20 25 30 35 40 45-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Time (h)
Co
ntr
ol In
put
(kW
) | E
nerg
y (
kW
h)
Control Input U (Case 1)
Control Input U (Case 2)
Energy (Case 1)
Energy (Case 2)
5 10 15 20 25 30 35 40 4518
19
20
21
22
23
24
25
26
Time (h)
Tem
pera
ture
(ºC
)
Temp. Ref.
Generated Ref.
Indoor Temp.
MPC AND PI CONTROL IN THERMAL COMFORT SYSTEMS
80
Figure 4.11. Power and energy consumption with limited power.
In Figure 4.10 can be seen that due the constrained power, Figure 4.11, the indoor temperature
cannot always follow the reference. The amount of energy expended grows a bit comparatively
with Figure 4.9.
4.3.4 Conclusions
In this chapter, in order to provide thermal house comfort, a MPC control technique associated
with a inner loop PI was presented.
It could be observed through the simulations and analysis results that good performances were
obtain by both approaches (with and without MPC). In general the error and the energy
consumption were considerably reduced with the MPC implementation.
In particular, results show that in the “comfort zone” MPC controller reduces considerably the
energy consumption with less power “peaks”.
The MPC power constrained shows that the approach can find a fair trade-off due to the
combination of the anticipative effect with available power, in order to maintain the temperature
near to the comfort zone, even when indoor temperature cannot follow the desire reference.
5 10 15 20 25 30 35 40 45-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
1.2
1.4
Time (h)
Contr
ol In
put
(kW
) | E
nerg
y (
kW
h)
Control Input U
Umax
Umin
Energy
81
Chapter 5
5 Thermal Comfort with Demand Side
Management Using Distributed MPC
5.1 Introduction
Emerging new technologies like distributed generation, distributed storage, and demand-side
load management will change the way we consume and produce energy. These techniques
enable the possibility to reduce the greenhouse effects and improve grid stability, balancing the
demand and supply. Also, the reduction in CO2 emissions and the introduction of renewable
sources based generation have become important topics today. However, these renewable
resources are mainly given by very fluctuating and uncontrollable sun, water, and wind power.
The work here presented intends to support the expected introduction of a large penetration
level of renewable sources, in particularly providing a solution for implementation in an
environment where buildings are mainly supplied by this type of resource. Especially because,
reducing energy consumption in buildings is a trend in the world today due to economic aspects
or environmental reasons. This chapter presents a DMPC for indoor thermal comfort which
simultaneously optimizes the consumption of a limited shared energy resource. The control
objective of each subsystem is to minimize the heating/cooling energy cost while maintaining
the indoor temperature and consumed power inside bounds. In a distributed coordinated
environment, the control uses multiple dynamically coupled agents (one for each
subsystem/house) aiming to achieve satisfaction of coupling constraints. In accordance with the
hourly power demand profile, each house assigns a priority level indicating how much is willing
to bid in auction for consumption of limited clean resource. This procedure allows the hourly
variation of bidding values and, consequently, agent’s orders for accessing clean energy also
THERMAL COMFORT WITH DEMAND SIDE MANAGEMENT USING DISTRIBUTED MPC
82
varies. Despite of power constraints, all houses also have thermal comfort constraints that must
be fulfilled. The system is simulated with several houses in a distributed environment.
5.2 Implemented global scenario
The scenario considers two types of available energy resources, the green and the red. The
green or clean resource must always be consumed (it is not disposable), it is limited to a
maximum available value and it is considered a time variable resource. In opposition, the red is
always available and it is considered a dirty resource, more expensive than the green. Therefore,
if the green resource is insufficient to satisfy the house’s required demand, the red must be
consumed with an increase in the total energy cost. To encourage clean resource consumption, it
is considered that green energy price per kWh has a maximum auction value and it is always
cheaper than the red energy price. The agents (one by each house) must bid in an auction
(provided by the MO), the price that they are willing to pay to consume the green resource. The
agents make their bid in the auction with one day ahead to show how much is intended to pay
per kWh to consume green resource in each one of the next 24 hours. The next dots summarize
all assumptions made:
Divisions must share a limited predictable renewable green resource;
The green or clean resource must always be consumed or stored and it is limited to
a maximum available that varies in time.
Agents make their bid in the auction with one day ahead to show how much is
intended to pay per kWh for consumption of green resource, in each one of the
next 24 hours;
Access to green resource is done hourly in accordance to the made bid value. The
one that is able to pay more uses the needed stock first, and the second can use only
the remaining energy, and so on;
Outside temperatures, disturbances and daily comfort temperature bounds are
known by each system inside the predictive horizon (HP);
If the green resource is insufficient to satisfy the demand required by the
houses/divisions, the red resource (from grid) must be consumed;
The red resource has a fixed kW price that is always higher than the green to
promote the renewable energy consumption;
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
83
The consumed red resource for comfort purposes implies a penalty in the final cost
function (5.1) due to the soft constraint violation, which is imposed by the
maximum available green resource if exceeded.
The MPC optimization is solved by each TCA at each time step according to the sequential
DMPC scheme depicted in Figure 5.1. The green available power forecast is received by the
agent who made the highest bid (first in the sequential access scheme) and this information is
used as power constraint value in (5.10). The agent optimization problem predicts the
consumption, subtracts it to the initially received available maximum and passes the information
to the next on the sequence list.
MPC 1
MPC 2
TCA 1
TCA 2
MPC NW -1 TCA NS -1
MPC NW TCA NS
uTgreen
*12 1TgreenU u u
*11 1
S
S S
NN NU U u
*iT
*iu
*23 3 1U U u
21u
11u
1
1SN
u
11v
21v
1
1SN
v
11T
21T
1
1SN
T
1SN
u
1SN
v
1SN
T
Figure 5.1. Implemented sequential architecture scheme.
In Figure 5.1, 𝑢𝑇𝑔𝑟𝑒𝑒𝑛 represents the green resource forecasts, 𝑈𝑖 the maximum available
expected green resource for TCA (i), 𝑇𝑖 and 𝑇𝑖∗ are the indoor temperature and indoor
temperature prediction, 𝑢𝑖 and 𝑢𝑖∗ are the optimal power generated by the MPC controller and
the predicted consumption and 𝑣𝑖 represents the disturbances.
THERMAL COMFORT WITH DEMAND SIDE MANAGEMENT USING DISTRIBUTED MPC
84
5.3 MPC formalization
At each time step, each one of the agents must solve his MPC problem. The objectives are:
minimize the energy consumption to heating and cooling; minimize the peak power
consumption; maintain the zones within a desired temperature range and maintain the used
power within the green available bounds. Feedback stability is provided by choosing a
sufficiently long predictive horizon and proven by results presented in next shapters. Feasibility
is achieved by the use of soft constraints in the optimization problem formulation as explained
in the sequel.
The generic optimization problem to be solved by each agent at each instant, assumes the
following form:
𝑚𝑖𝑛𝑈,��,𝜀,��,𝛾
𝐽𝑖(𝑘) = ∑ [∑𝑢𝑙𝑖(𝑘 + 𝑗)𝑇𝜑𝑙
𝑖𝑢𝑙𝑖(𝑘 + 𝑗)⏟
𝐶𝑜𝑛𝑠𝑢𝑚𝑝𝑡𝑖𝑜𝑛
𝑁𝑑𝑖
𝑙=1
]
𝐻𝑝−1
𝑗=0
+∑𝜙𝑖𝑚𝑎𝑥 {𝑢𝑙𝑖2(𝑘), . . . . , 𝑢𝑙
𝑖2(𝑘 + 𝐻𝑝 − 1)}⏟ 𝑃𝑜𝑤𝑒𝑟 𝑝𝑒𝑎𝑘𝑠
𝑁𝑑𝑖
𝑙=1
+∑[∑(𝜀𝑙𝑖(𝑘 + 𝑗)𝑇𝛯𝑙
𝑖𝜀𝑙𝑖(𝑘 + 𝑗⏟
𝐶𝑜𝑚𝑓𝑜𝑟𝑡 𝑣𝑖𝑜𝑙𝑎𝑡𝑖𝑜𝑛
) + 𝛾𝑙𝑖(𝑘 + 𝑗)𝑇𝛹𝑙
𝑖𝛾𝑙𝑖(𝑘 + 𝑗)⏟
𝑃𝑜𝑤𝑒𝑟 𝑣𝑖𝑜𝑙𝑎𝑡𝑖𝑜𝑛
)
𝑁𝑑𝑖
𝑙=1
]
𝐻𝑝
𝑗=1
,
(5.1)
𝑚𝑖𝑛𝑈𝑖,𝜀𝑖,𝛾𝑖
𝐽𝑖(𝑘) = 𝜀𝑖𝑇(𝑘)𝛯𝑖𝜀𝑖(𝑘) + 𝛾𝑖
𝑇(𝑘)𝛹𝑖𝛾𝑖(𝑘) + 𝑈𝑖𝑇(𝑘)𝑅𝑖𝑈𝑖(𝑘)
+∑𝜙𝑖𝑚𝑎𝑥 {𝑢𝑙𝑖2(𝑘), . . . . , 𝑢𝑙
𝑖2(𝑘 + 𝐻𝑝 − 1)}
𝑁𝑑𝑖
𝑙=1
, (5.2)
with,
𝑈𝑖(𝑘) =
[ 𝑈𝑖1(𝑘)
𝑈𝑖2(𝑘)⋮𝑈𝑖𝑁𝑑(𝑘)]
, (5.3)
𝜀𝑖(𝑘) =
[ 𝜀𝑖1(𝑘)
𝜀𝑖2(𝑘)
⋮
𝜀𝑖𝑁𝑑(𝑘)
𝜀𝑖1(𝑘)
𝜀𝑖2(𝑘)
⋮𝜀𝑖𝑁𝑑(𝑘)]
, 𝜀𝑖𝑙(𝑘) =
[ 𝜀𝑖𝑙(𝑘 + 1)
𝜀𝑖𝑙(𝑘 + 2)
⋮
𝜀𝑖𝑙(𝑘 + 𝐻𝑃)]
, 𝜀𝑖𝑙(𝑘) =
[ 𝜀𝑖𝑙(𝑘 + 1)
𝜀𝑖𝑙(𝑘 + 2)
⋮𝜀𝑖𝑙(𝑘 + 𝐻𝑃)]
, (5.4)
𝛾𝑖(𝑘) = [𝛾𝑖(𝑘)
𝛾𝑖(𝑘)], 𝛾
𝑖(𝑘) =
[ 𝛾𝑖(𝑘 + 1)
𝛾𝑖(𝑘 + 2)
⋮𝛾𝑖(𝑘 + 𝐻𝑃)]
, 𝛾𝑖(𝑘) =
[ 𝛾𝑖(𝑘 + 1)
𝛾𝑖(𝑘 + 2)
⋮𝛾𝑖(𝑘 + 𝐻𝑃)]
. (5.5)
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
85
Resulting in quadratic optimization problem in the compact form
𝑚𝑖𝑛 𝑍𝑖𝐽𝑖(𝑘) = 𝑍𝑖
𝑇𝛩𝑍𝑖 +∑𝜙𝑖𝑙𝑚𝑎𝑥 {𝑢𝑙𝑖2(𝑘), . . . . , 𝑢𝑙
𝑖2(𝑘 + 𝐻𝑝 − 1)}
𝑁𝑑𝑖
𝑙=1
. (5.6)
With
𝑍𝑖 = [𝑈𝑖𝜀𝑖𝛾𝑖
], 𝛩 = [
𝜑𝑖 0 00 𝛯𝑖 00 0 𝛹𝑖
], (5.7)
and subject to the following constraints,
𝑥𝑙𝑖(𝑘 + 𝑗 + 1) = 𝐴𝑙𝑙
𝑖𝑖𝑥𝑙𝑖(𝑘 + 𝑗) + 𝐵𝑙
𝑖𝑢𝑙𝑖(𝑘 + 𝑗) + ∑ (𝐴𝑙𝑔
𝑖𝑖 ��𝑙𝑖(𝑘 + 𝑗))
𝑁𝑑𝑖
𝑔=1(𝑔≠𝑙)⏟ 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑 𝑡𝑒𝑚𝑝𝑒𝑟𝑎𝑡𝑢𝑟𝑒𝑠 𝑓𝑟𝑜𝑚 𝑎𝑑𝑗𝑎𝑐𝑒𝑛𝑡 𝑎𝑟𝑒𝑎𝑠
𝑖𝑛𝑠𝑖𝑑𝑒 𝑡ℎ𝑒 𝑠𝑎𝑚𝑒 ℎ𝑜𝑢𝑠𝑒
+ ∑ ∑ (𝐴𝑙𝑚𝑖ℎ ��𝑚
ℎ (𝑘 + 𝑗))
𝑁𝑑ℎ
𝑚=1
𝑁𝑠
ℎ=1(ℎ≠𝑖)⏟ 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑 𝑡𝑒𝑚𝑝𝑒𝑟𝑎𝑡𝑢𝑟𝑒𝑠 𝑓𝑟𝑜𝑚 𝑎𝑑𝑗𝑎𝑐𝑒𝑛𝑡 𝑎𝑟𝑒𝑎𝑠
𝑓𝑟𝑜𝑚 𝑜𝑡ℎ𝑒𝑟 ℎ𝑜𝑢𝑠𝑒𝑠
+ 𝑣𝑙𝑖(𝑘 + 𝑗),
, 𝑗 = 1. . . 𝐻𝑃
(5.8)
𝑇𝑙𝑖(𝑘 + 𝑗) − 𝜀𝑙
𝑖(𝑘 + 𝑗) ≤ 𝑥𝑙𝑖(𝑘 + 𝑗) ≤ ��𝑙
𝑖(𝑘 + 𝑗) + 𝜀��𝑖(𝑘 + 𝑗), (5.9)
𝑈𝑖(𝑘 + 𝑗 − 1) − 𝛾𝑖(𝑘 + 𝑗 − 1) ≤∑𝑢𝑙𝑖(𝑘 + 𝑗 − 1)
𝑁𝑑𝑖
𝑙=1
≤ 𝑈𝑖(𝑘 + 𝑗 − 1) + ��𝑖(𝑘 + 𝑗 − 1), (5.10)
𝛾𝑖 , ��𝑖 , 𝜀𝑙𝑖, 𝜀��
𝑖 ≥ 0. (5.11)
In (5.1) 𝑁𝑑𝑖is the number of divisions of house (i), 𝑢𝑖𝑙 represents the power control inputs from
house (i) division (l), 𝜙𝑖is the penalty on peak power consumption, 𝛯𝑖 is the penalty on the
comfort constraint violation, Ψithe penalty on the power constraint violation and HP is the
length of the prediction horizon. In (5.9), 𝜀��𝑖 and 𝜀𝑙
𝑖 are the vectors of temperature violations that
are above and below the desired comfort zone defined by ��𝑙𝑖 and 𝑇𝑙
𝑖. In (5.10), the coupled
power constraint, 𝛾𝑖 and 𝛾𝑖are the power violations that are above or lower the maximum,𝑈𝑖,
and minimum, 𝑈𝑖, available green power for heating/cooling the house, with 𝑈𝑖 = −𝑈𝑖. Remark
that, in each TCA (i), the power sum in all divisions cannot exceed 𝑈𝑖 and that in (5.8) 𝑥𝑙𝑖∗and
𝑢𝑙𝑖∗represent the predicted temperature and power consumption within the prediction horizon.
With this approach each TCA can have different hourly penalties, allowing the consumer to
choose between more/less comfort and cost during the day as Figure 5.2 exemplify. This
procedure was implemented in Algorithm II.
THERMAL COMFORT WITH DEMAND SIDE MANAGEMENT USING DISTRIBUTED MPC
86
Figure 5.2. Daily consumer profile.
An example of formulization is made for and distributed situation where two distinct TCA’s are
dynamically and constraints coupled.
1 2
1u
1y2u
2y
Figure 5.3. Thermally coupled TCA’s.
For TCA1 and TCA2 it can be written
{
𝐶1𝑒𝑞
1 ��11 = 𝑢1
1 +𝑇𝑜𝑎 − 𝑇1
1
𝑅1𝑒𝑞1 +
𝑇12 − 𝑇1
1
𝑅11𝑒𝑞12 + 𝑃1𝑃𝑑
1 ,
𝐶1𝑒𝑞2 ��1
2 = 𝑢12 +
𝑇𝑜𝑎 − 𝑇12
𝑅1𝑒𝑞2 +
𝑇11 − 𝑇1
2
𝑅11𝑒𝑞12 + 𝑃1𝑃𝑑
2
(5.12)
{
𝐶1𝑒𝑞
1 ��11 = 𝑢1
1 − (1
𝑅1𝑒𝑞1 +
1
𝑅11𝑒𝑞12 )𝑇1
1 +𝑇𝑜𝑎 − 𝑇1
1
𝑅1𝑒𝑞1 +
𝑇12
𝑅11𝑒𝑞12 +
𝑇𝑜𝑎
𝑅1𝑒𝑞1 +𝑃1𝑃𝑑
1
𝐶1𝑒𝑞2 ��1
2 = 𝑢12 − (
1
𝑅1𝑒𝑞2 +
1
𝑅11𝑒𝑞12 )𝑇1
2 +𝑇𝑜𝑎 − 𝑇1
2
𝑅1𝑒𝑞2 +
𝑇11
𝑅11𝑒𝑞12 +
𝑇𝑜𝑎
𝑅1𝑒𝑞2 +𝑃1𝑃𝑑
2
(5.13)
[��11
��12] =
[ −
𝑅1𝑒𝑞1 + 𝑅11𝑒𝑞
12
𝑅1𝑒𝑞1 𝐶1𝑒𝑞
1 𝑅11𝑒𝑞12
1
𝐶1𝑒𝑞1 𝑅11𝑒𝑞
12
1
𝐶1𝑒𝑞2 𝑅11𝑒𝑞
12 −𝑅1𝑒𝑞2 + 𝑅11𝑒𝑞
12
𝑅1𝑒𝑞2 𝐶1𝑒𝑞
2 𝑅11𝑒𝑞12
]
[𝑇11
𝑇12] +
[ 1
𝐶1𝑒𝑞1 0
01
𝐶1𝑒𝑞2]
[𝑢11
𝑢12] +
[ 𝑇𝑜𝑎
𝐶1𝑒𝑞1 𝑅1𝑒𝑞
1 +𝑃1𝑃𝑑1
𝐶1𝑒𝑞1
𝑇𝑜𝑎
𝐶1𝑒𝑞2 𝑅1𝑒𝑞
2 +𝑃1𝑃𝑑2
𝐶1𝑒𝑞2]
. (5.14)
For TCA 1 the plant model representation (5.14) can be rewritten and changed into a discrete
model using Euler discretization with a sampling time of Δt.
��11 =
𝑇(𝑘 + 1) − 𝑇(𝑘)
∆𝑡, (5.15)
00:00 23:0001:00 02:00 03:00 04:00 05:00 06:00 07:00 08:00 09:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 17:00 18:00 19:00 20:00 21:00 22:00
ComfortConsumption Consumption Comfort Consumption Comfort
MAX - 23 ºC
MIN - 21 ºC
MAX - 22 ºC
MIN - 20 ºC
MAX - 21 ºC
MIN - 20 ºC
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
87
𝑇(𝑘 + 1) − 𝑇(𝑘)
∆𝑡= −
𝑅1𝑒𝑞1 + 𝑅11𝑒𝑞
12
𝑅1𝑒𝑞1 𝐶1𝑒𝑞
1 𝑅11𝑒𝑞12 ��1
1(𝑘) −1
𝐶1𝑒𝑞1 𝑅11𝑒𝑞
12 𝑇12(𝑘) +
1
𝐶1𝑒𝑞1 𝑢1
1(𝑘) +1
𝐶1𝑒𝑞1 𝑅1𝑒𝑞
1 𝑇𝑜𝑎(𝑘)
+1
𝐶1𝑒𝑞1 𝑃1𝑃𝑑
1 (𝑘), (5.16)
𝑇11(𝑘 + 1) = 𝐴11
11𝑇11(𝑘) + 𝐵1
1𝑢11(𝑘) + 𝐴11
12𝑇12(𝑘) + 𝑣1
1(𝑘), (5.17)
where
𝐴1 = 𝐴1111 = (1 −
𝑅1𝑒𝑞1 + 𝑅11𝑒𝑞
12
𝐶1𝑒𝑞1 𝑅1𝑒𝑞
1 𝑅11𝑒𝑞12 ∆𝑡), (5.18)
𝐷1 = 𝑣11 =
𝑃1𝑃𝑑1
𝐶1𝑒𝑞1 Δt +
Toa
𝑅1𝑒𝑞1 𝐶1𝑒𝑞
1 Δt,
𝐵1 = 𝐵11 =
1
𝐶1𝑒𝑞1 𝛥𝑡,
𝐶1 = 𝐴1112 =
1
𝑅11𝑒𝑞12 𝐶1𝑒𝑞
1 Δt.
For several instants, (5.17) can be written as follows,
𝑇11(1) = 𝐴1𝑇1
1(0) + 𝐵1𝑢11(0) + 𝐶1𝑇1
2(0) + 𝐷1(0), (5.19)
𝑇11(2) = (𝐴1)
2𝑇11(0) + 𝐴1𝐵1𝑢1
1(0) + 𝐴1𝐶1𝑇12(0) + 𝐴1𝐷1(0) + 𝐵1𝑢1
1(1) + 𝐶1𝑇12(1) + 𝐷1(1), (5.20)
𝑇11(3) = (𝐴1)
3𝑇11(0) + (𝐴1)
2𝐵1𝑢11(0) + (𝐴1)
2𝐶1𝑇12(0) + (𝐴1)
2𝐷1(0) + 𝐴1𝐵1𝑢11(1)
+ 𝐴1𝐶1𝑇12(1) + 𝐴1𝐷1(1) + 𝐵1𝑢1
1(2) + 𝐶1𝑇12(2) + 𝐷1(2), (5.21)
Therefore,
𝑀𝑎𝑡𝑈 =
[ 𝐵1 0 ⋯ ⋯ ⋯ 0𝐴1𝐵1 𝐵1 ⋱ ⋯ ⋯ ⋮
𝐴12𝐵1 𝐴1𝐵 𝐵1 ⋱ ⋯ ⋮
𝐴13𝐵1 𝐴1
2𝐵1 𝐴1𝐵1 𝐵1 ⋱ ⋮⋮ ⋮ ⋮ ⋮ 𝐵1 0
𝐴1𝐻𝑃𝐵1 𝐴1
𝐻𝑃−1𝐵1 ⋯ ⋯ 𝐴1𝐵1 𝐵1]
, (5.22)
THERMAL COMFORT WITH DEMAND SIDE MANAGEMENT USING DISTRIBUTED MPC
88
𝑀𝑎𝑡𝑇𝑥 =
[ 𝐶1 0 ⋯ ⋯ ⋯ 0𝐴1𝐶1 𝐶1 ⋱ ⋯ ⋯ ⋮
𝐴12𝐶1 𝐴1𝐶1 𝐶1 ⋱ ⋯ ⋮
𝐴13𝐶1 𝐴1
2𝐶1 𝐴1𝐶1 𝐶1 ⋱ ⋮⋮ ⋮ ⋮ ⋮ 𝐶1 0
𝐴1𝐻𝑃𝐶1 𝐴1
𝐻𝑃−1𝐶1 ⋯ ⋯ 𝐴1𝐶1 𝐶1]
, (5.23)
𝑀𝑎𝑡𝐷 =
[ 1 0 ⋯ ⋯ ⋯ 0𝐴1 1 ⋱ ⋯ ⋯ ⋮
𝐴12 𝐴1 1 ⋱ ⋯ ⋮
𝐴13 𝐴1
2 𝐴1 1 ⋱ ⋮⋮ ⋮ ⋮ ⋮ 1 0𝐴1
𝐻𝑃 𝐴1𝐻𝑃−1 ⋯ ⋯ 𝐴1 1]
, (5.24)
𝑀𝑎𝑡𝑇 = [𝐴1 𝐴12 𝐴1
3 ⋯ 𝐴1𝐻𝑃].
(5.25)
It is considered that the constraints (5.9) and (5.10) must assume the form of Ax<B with,
𝑥 = [𝑢11; 𝜀1; 𝜀1; 𝛾1; ��1]
𝑇
,
(5.26)
where, ε1 and ε1 are the vectors of temperature violations that are above and below the desired
comfort zone defined by 𝑇11 and 𝑇1
1, and 𝛾1 and 𝛾1 are the power violations that are above or
lower the maximum.
Temperature constraints: 𝑇11(𝑘 + 1) = 𝐴1𝑇1
1(𝑘) + 𝐵1𝑢11(𝑘) + 𝐶1𝑇1
2(𝑘) + 𝐷1(𝑘),
𝑇11 − 𝜀1 ≤ 𝑇1
1 ≤ 𝑇11, +𝜀1,
(5.27)
𝑇11 −
[ 𝜀1(1)
𝜀1(2)
⋮⋮⋮𝜀1(𝐻𝑃)]
≤
[ 𝐴1𝐴1
2
𝐴13
⋮⋮𝐴1
𝐻𝑃]
𝑇11(0)
[ 𝐵1 0 ⋯ ⋯ ⋯ 0𝐴1𝐵1 𝐵1 ⋱ ⋯ ⋯ ⋮
𝐴12𝐵1 𝐴1𝐵 𝐵1 ⋱ ⋯ ⋮
𝐴13𝐵1 𝐴1
2𝐵1 𝐴1𝐵1 𝐵1 ⋱ ⋮⋮ ⋮ ⋮ ⋮ 𝐵1 0
𝐴1𝐻𝑃𝐵1 𝐴1
𝐻𝑃−1𝐵1 ⋯ ⋯ 𝐴1𝐵1 𝐵1]
[ 𝑢11(0)
𝑢11(1)⋮⋮⋮𝑢11(𝐻𝑃 − 1)]
+
[ 𝐶1 0 ⋯ ⋯ ⋯ 0𝐴1𝐶1 𝐶1 ⋱ ⋯ ⋯ ⋮
𝐴12𝐶1 𝐴1𝐶1 𝐶1 ⋱ ⋯ ⋮
𝐴13𝐶1 𝐴1
2𝐶1 𝐴1𝐶1 𝐶1 ⋱ ⋮⋮ ⋮ ⋮ ⋮ 𝐶1 0
𝐴1𝐻𝑃𝐶1 𝐴1
𝐻𝑃−1𝐶1 ⋯ ⋯ 𝐴1𝐶1 𝐶1]
[ 𝑇12(0)
𝑇12(1)⋮⋮⋮𝑇12(𝐻𝑃 − 1)]
+
(5.28)
[ 1 0 ⋯ ⋯ ⋯ 0𝐴1 1 ⋱ ⋯ ⋯ ⋮
𝐴12 𝐴1 1 ⋱ ⋯ ⋮
𝐴13 𝐴1
2 𝐴1 1 ⋱ ⋮⋮ ⋮ ⋮ ⋮ 1 0𝐴1
𝐻𝑃 𝐴1𝐻𝑃−1 ⋯ ⋯ 𝐴1 1]
[ 𝐷1(0)
𝐷1(1)⋮⋮⋮𝐷1(𝐻𝑃 − 1)]
≤ 𝑇11, +
[ 𝜀1(1)
𝜀1(2)⋮⋮⋮𝜀1(𝐻𝑃)]
.
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
89
Lower temperature constraint bound
𝑇11 − 𝜀1 ≤ 𝑇1
1, (5.29)
𝑇11 − 𝜀1 ≤ 𝑀𝑎𝑡𝑇 × 𝑇1
1 +𝑀𝑎𝑡𝑈 × 𝑢11 +𝑀𝑎𝑡𝑇𝑥 × 𝑇1
2 +𝑀𝑎𝑡𝐷 × 𝐷1, (5.30)
−(𝜀 +𝑀𝑎𝑡𝑈 × 𝑢11
⏟ 𝐴1
) ≤ −(𝑇11 −𝑀𝑎𝑡𝑇 × 𝑇1
1 +𝑀𝑎𝑡𝑇𝑥 × 𝑇12 +𝑀𝑎𝑡𝐷 × 𝐷1⏟
𝐵1
), (5.31)
𝐴1 = [𝑀𝑎𝑡𝑈𝐻𝑃×𝐻𝑃 𝐼𝐻𝑃×𝐻𝑃 0𝐻𝑃×𝐻𝑃 0𝐻𝑃×𝐻𝑃 0𝐻𝑃×𝐻𝑃]. (5.32)
Similarly, for the upper temperature bound can be written
𝑇11 ≤ 𝑇1
1+ 𝜀1, (5.33)
𝑀𝑎𝑡𝑇 × 𝑇11 +𝑀𝑎𝑡𝑈 × 𝑢1
1 +𝑀𝑎𝑡𝑇𝑥 × 𝑇12 +𝑀𝑎𝑡𝐷 × 𝐷1 ≤ 𝑇1
1+ 𝜀1, (5.34)
(𝜀1 +𝑀𝑎𝑡𝑈 × 𝑢11⏟
𝐴2
) ≤ −(𝑇11+𝑀𝑎𝑡𝑇 × 𝑇1
1 +𝑀𝑎𝑡𝑇𝑥 × 𝑇12 +𝑀𝑎𝑡𝐷 × 𝐷1⏟
𝐵2
), (5.35)
𝐴2 = [𝑀𝑎𝑡𝑈𝐻𝑃×𝐻𝑃 0𝐻𝑃×𝐻𝑃 − 𝐼𝐻𝑃×𝐻𝑃 0𝐻𝑃×𝐻𝑃 0𝐻𝑃×𝐻𝑃]. (5.36)
Power constraints also assume the form Ax<B, so to the lower bound,
𝑈1 − 𝛾1 ≤ 𝑈1, (5.37)
−𝛾1 − 𝑈1⏟ 𝐴3
≤ −𝑈1⏟𝐵3
, (5.38)
𝐴3 = [−𝐼𝐻𝑃×𝐻𝑃 0𝐻𝑃×𝐻𝑃 0𝐻𝑃×𝐻𝑃 − 𝐼𝐻𝑃×𝐻𝑃 0𝐻𝑃×𝐻𝑃]. (5.39)
Upper power constraint bound
𝑈1 ≤ ��1 + γ1, (5.40)
𝑈1 − γ1⏟ 𝐴4
≤ ��1⏟𝐵4
, (5.41)
A4 = [𝐼𝐻𝑃 ×𝐻𝑃 0𝐻𝑃×𝐻𝑃 0𝐻𝑃×𝐻𝑃 0𝐻𝑃×𝐻𝑃 − 𝐼𝐻𝑃×𝐻𝑃]. (5.42)
Finally for all soft constrains in the form of Ax<B results in,
[
−A1A2A3A4
] [𝑢11; ε1; ε1; γ1; γ1]
𝑇= [
−𝐵1𝑇
𝐵2𝑇
𝐵3𝑇
𝐵4𝑇
].
(5.43)
THERMAL COMFORT WITH DEMAND SIDE MANAGEMENT USING DISTRIBUTED MPC
90
5.3.1 Algorithm I - Implemented sequential scheme
After the methodology description, the algorithm is described. As mentioned it is considered
that the sequence order was already stablished by a previous auction. The hourly access
sequence is storage in AO(NS×HP). Remark that, only the divisions that thermally interact pass to
each other the information about future indoor temperatures prevision.
Algorithm I - DSM-DMPC Implemented sequential scheme
Required for all TCA’s: Thermal disturbances, comfort temperature gap, hourly constraints
parameters and auction bid.
For all TCA’s Wi initialize:
BV bid value inside the HP PH1
for k=1 to NC
for i=1 to NS(the number of TCA’s)
Apply to all TCA 𝑢𝑙𝑖 (𝑘 − 1: 𝑘 − 1 + 𝐻𝑃) from k-1 instant to obtain 𝑇𝑙
𝑖(𝑘: 𝑘 +𝐻𝑃)
Communicate temperature predictions to adjacent TCA’s
Update available green resource
Calculate the optimal control sequence 𝑢𝑙𝑖(1:HP) solving OPi (5.1)-(5.11) with power constraints
(5.10) given by 𝑈𝑖 and ��𝑖
Update available green resource values for next TCA
��𝑖+1(𝑘: 𝑘 + 𝐻𝑃) = ��𝑖(𝑘: 𝑘 + 𝐻𝑃) −∑𝑢𝑙𝑖(𝑘: 𝑘 + 𝐻𝑃)
𝑁𝑑𝑖
𝑙=1
end for
end for
Remark: generically, 𝑋(𝑘: 𝑘 + 𝐻𝑃) represents a line vector (1 × 𝐻𝑃) containing values from
𝑥(𝑘) to 𝑥(𝑘 + 𝐻𝑃), and 𝑌(𝑝, 𝑘: 𝑘 + 𝐻𝑃), represents the line p of a matrix (𝑃 × 𝐻𝑃) containing
values from y(𝑝, 𝑘) to y(𝑝, 𝑘 + 𝐻𝑃).
91
Chapter 6
6 DMPC for Thermal House Comfort
with Sequential Access Auction
6.1 Introduction
This chapter presents, in a model predictive control multi-agent systems context, an integrative
methodology to manage energy networks from the demand side with strong presence of
intermittent energy sources and with energy storage in house-hold or car batteries. In particular
is presented a distributed model predictive control solution for indoor comfort that
simultaneously optimizes the usage of a limited shared energy resource via a demand side
management perspective (Barata et al., 2012b). The control is applied individually to a set of
Thermal Control Areas, demand units, where the objective is to minimize the energy cost while
maintaining the indoor temperature inside a comfort zone, without exceeding the limited and
shared energy resource. The Thermal Control Areas are generally thermodynamically connected
in the distributed environment and also are coupled by energy related constraints. The overall
system predicts indoor environmental conditions in buildings with different plans which are
modelled using an electro-thermal modular scheme (Chapter 3.4). For control purposes, this
modular scheme allows an easy modelling of buildings with different plans where adjacent
areas can thermally interact.
The results are presented with two different approaches. In Chapter 6.2, the energy split
performance is based on a fixed sequential order, established from a previously done auction
wherein the bids are made by each Thermal Control Area, acting as demand side management
agents and based on the energy daily price. In Chapter 6.3, bids can be made in accordance to
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
92
the consumption needs. Each TCA has an hourly known consumption profile and the bid value
is directly related to that consumption behaviour. Thus, to the assumptions made in Chaper 5.2
the next dots are added, stipulating the incremented specificities made in this chapter:
Green resource which is not consumed at the final of a certain instant is stored in
batteries up to capacity value (BcV). When BcV is reached it is considered that the
remaining green resource is delivered to the grid;
The access order to green resource varies hourly;
Each TCA has a known fixed 24 hours consumption profile, 𝐶𝑤𝑙𝑖(𝑘: 𝑘 + 𝐻𝑃) and it
is established a priority level from 1 (low) to 3 (high) for each hour indicating how
important it is to have available resource to supply the load;
The bid value of each house is made in accordance with the chosen priority level,
the hours with higher priority levels indicate high consumption and consequently a
higher bid value;
The access to green resource is done hourly in accordance with the made bid value.
The one that is able to pay more uses the needed stock first and the second can use
only the remaining energy, and so on.
The cost function penalty values may hourly vary;
The developed solutions are explained by Algorithm II and are applied to different scenarios
wherein the results illustrate the benefits of the proposed approach.
6.2 Access order with fixed stablished sequence
To explain and prove the concept, simulation results were made in accordance with the scheme
presented in Figure 6.2.
Figure 6.1: Heat transfer between divisions example.
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
93
Figure 6.2: System implementation scheme block diagram.
Three houses are considered, two of them thermally interacting (with a thermal resistance
between them of R12=30ºC/kW as shown in Figure 6.1 and the third is isolated.
6.2.1 Results
It is considered that all TCA´s have the same outdoor temperature presented in Figure 6.3.
Figure 6.3. Outdoor temperature forecasting.
The thermal characteristics and prices for all agents and scenarios are presented in Table 6.1.
House 3
House 1
SYSTEM DYNAMICALLY COUPLED
AND COUPLED BY CONSTRAINTS SYSTEM ONLY COUPLED BY CONSTRAINTS
House 2
0 5 10 15 2015
20
25
30
35
40
Time (h)
Te
mp
era
ture
(ºC
)
Temp. Outside
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
94
Table 6.1. Distributed parameters
Parameter A1 A2 A3 Units
Req 50 50 75 ºC/kW
Ceq 9.2103 9.2103 9.2103 kJ/ºC
Green Price (per kWh) 0.09 0.08 0.07 €
Red Price(per kWh) 0.18 €
As mentioned, agents can also have distinct penalties on power and temperature constraints
violations, they can hourly privilege comfort or cost according to consumer choice. Therefore,
to explore the concept several scenarios are here presented.
6.2.1.1 Scenario I - Balanced parameterization
The first, it is considered a balanced scenario, with no energy storage and no explicit preference
between comfort or consumption. Table 6.2 shows the penalty values that are used in the first
scenario.
Table 6.2. Scenario I - Penalty values
Parameter A1 A2 A3
𝛯 50 50 50
100 100 30
ϕ 2 2 2
1 1 1
The thermal disturbances forecasts and the indoor temperature with its constraints for the TCA,
A1, A2 and A3 have the profile presented in Figure 6.4, Figure 6.5 and Figure 6.6 respectively.
Figure 6.4. Scenario I - Disturbance forecasting and indoor temperature A1.
0 5 10 15 200
0.5
1
1.5
2
Pow
er
(kW
)
0 5 10 15 2018
19
20
21
22
23
Time (h)
Tem
pera
ture
(ºC
)
Indoor Temp. (ºC) Temp. Max (ºC) Temp. Min (ºC)
Disturbance Load A1 (kW)
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
95
Figure 6.5. Scenario I - Disturbance forecasting and indoor temperature A2.
Figure 6.6. Scenario I - Disturbance forecasting and indoor temperature A3.
Figure 6.7. Scenario I - A1 power profile.
0 5 10 15 2021
21.5
22
22.5
23
23.5
24
Time (h)
Tem
pera
ture
(ºC
)
0 5 10 15 200
0.5
1
1.5
2
Pow
er
(kW
)
Indoor Temp. (ºC) Temp. Max (ºC) Temp. Min (ºC)
Disturbance Load A2 (kW)
0 5 10 15 200
0.5
1
1.5
2
2.5
3
Pow
er
(kW
)
0 5 10 15 2020
21
22
23
24
25
Time (h)
Tem
pera
ture
(ºC
)
Disturbance Load A3 (kW)
Indoor Temp. (ºC) Temp. Max (ºC) Temp. Min (ºC)
0 5 10 15 20-2
-1
0
1
2
3
Pow
er
(kW
)
0 5 10 15 200
0.5
1
1.5
2
2.5
Time (h)
Pow
er
(kW
)
Power Max (kW) Power Min (kW) Consumption (kW)
Total Available (kW) Red Resource (kW) Green Resource (kW)
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
96
Despite the thermal disturbances, it can be seen in that the indoor temperatures are
predominantly inside their narrow bounds. The Figure 6.7, Figure 6.8 and Figure 6.9 show, for
TCA, A1, A2 and A3 respectively, the used power to acclimatize the spaces and the available
green resource to do it. The figures also show the amount and the type of the consumed power.
Figure 6.8. Scenario I - A2 power profile.
Figure 6.9. Scenario I - A3 power profile.
The power consumption behaviour shows that the controllers try to consume only when the
green resource is available. The compromise between comfort and consumption is notorious,
the controllers are always attempting to accomplish both constraints, but when is not possible it
compensates with one inside range and the other outside.
0 5 10 15 20-2
-1
0
1
2
Pow
er
(kW
)
0 5 10 15 200
0.5
1
1.5
2
Time (h)
Pow
er
(kW
)
Red Resource (kW) Green Resource (kW) Total Available (kW)
Power Max (kW) Consumption (kW) Power Min (kW)
0 5 10 15 20-2
-1
0
1
2
Pow
er
(kW
)
0 5 10 15 200
0.5
1
1.5
2
Time (h)
Pow
er
(kW
)
Consumption (kW) Power Max (kW) Power Min (kW)
Red Resource (kW) Green Resource (kW) Total Available (kW)
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
97
Figure 6.10. Scenario I - Global consumption characterization.
Figure 6.11. Scenario I - Power profile.
Figure 6.12. Scenario I - Heating/cooling total cost.
0 5 10 15 200
0.5
1
1.5
2
2.5
Time (h)
Po
we
r (k
W)
Total Available (kW)
Total Red (kW)
Total Green (kW)
0 5 10 15 200
5
10
15
Time (h)
Pow
er
(kW
)
House 1 (kW)
House 2 (kW)
House 3 (kW)
0 5 10 15 200
0.2
0.4
0.6
0.8
1
1.2
Time (h)
Cost
(€)
House 3
House 1
House 2
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
98
Being the last in the sequence, A3 receives merely the remainder resource, which in several
periods is null obliging to a major consumption of red resource. Due this situation, despite A3
having the lowest consumption, Figure 6.11, is the one that have the highest cost Figure 6.12.
6.2.1.2 Scenario II - Balanced parameterization with storage
The second scenario intends to show the benefit of having energy storage, been the only
difference with Scenario I the increment of a set of batteries with 1.5 kWh of capacity, Figure
6.22. The thermal disturbances forecasts and the indoor temperature with its constraints for the
TCA, A1, A2 and A3 in this Scenario II, have the profile presented in Figure 6.13, Figure 6.14
and Figure 6.15 respectively.
Figure 6.13. Scenario II - Disturbance forecasting and indoor temperature A1.
Figure 6.14. Scenario II - Disturbance forecasting and indoor temperature A2.
0 5 10 15 200
0.5
1
1.5
2
Pow
er
(kW
)
0 5 10 15 2018
19
20
21
22
23
24
Time (h)
Tem
pera
ture
(ºC
)
Disturbance Load A1(kW)
Indoor Temp. (ºC) Temp. Max (ºC) Temp. Min (ºC)
0 5 10 15 200
0.5
1
1.5
2
Pow
er
(kW
)
0 5 10 15 2021
21.5
22
22.5
23
23.5
24
Time (h)
Tem
pera
ture
(ºC
)
Disturbance Load A2(kW)
Indoor Temp. (ºC) Temp. Max (ºC) Temp. Min (ºC)
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
99
Figure 6.15. Scenario II - Disturbance forecasting and indoor temperature A3.
Figure 6.16. Scenario II – A1 power profile.
Figure 6.17. Scenario II – A2 power profile.
0 5 10 15 200
1
2
3
Pow
er
(kW
)
0 5 10 15 2020
21
22
23
24
25
Time (h)
Tem
pera
ture
(ºC
)
Disturbance Load A3(kW)
Indoor Temp. (ºC) Temp. Max (ºC) Temp. Min (ºC)
0 5 10 15 20
-4
-2
0
2
4
6
Pow
er
(kW
)
0 5 10 15 200
1
2
3
4
5
Time (h)
Pow
er
(kW
)
Consumption (kW) Power Max (kW) Power Min (kW)
Red Resource (kW) Green Resource (kW) Total Available (kW)
0 5 10 15 20
-4
-2
0
2
4
6
Pow
er
(kW
)
0 5 10 15 200
0.5
1
1.5
2
2.5
3
3.5
Time (h)
Pow
er
(kW
)
Red Resource (kW) Green Resource (kW) Total Available (kW)
Power Max (kW) Power Min (kW) Consumption (kW)
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
100
Figure 6.18. Scenario II – A3 power profile.
Despite the thermal disturbances, it can be seen in that the indoor temperatures are
predominantly inside their narrow bounds.
As expected, with this energy supplement the power constraint of all TCA’s was respected,
exceptionally in the interval between 12-15h in A3, where a small amount of red resource was
consumed.
Figure 6.19. Scenario II - Global consumption characterization.
Comparatively with Figure 6.10, Figure 6.19 shows that the red resource consumption was
significantly reduced. The agent that most benefit with this modification was A3, the energy
surplus allows it to consume almost exclusively green resource bought at lower bid price , and
thus, is the one that presents lesser costs, Figure 6.21.
0 5 10 15 20-4
-2
0
2
4
Pow
er
(kW
)
0 5 10 15 200
1
2
3
Time (h)
Pow
er
(kW
)
Power Max (kW) Consumption (kW) Power Min (kW)
Red Resource (kW) Green Resource (kW) Total Available (kW)
0 5 10 15 200
0.5
1
1.5
2
2.5
3
3.5
Time (h)
Po
we
r (k
W)
Total Available (kW)
Total Red (kW)
Total Green (kW)
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
101
Figure 6.20. Scenario II - Power profile.
Figure 6.21. Scenario II - Heating/cooling total cost.
Figure 6.22. Scenario II - Batteries profile.
0 5 10 15 200
2
4
6
8
10
12
14
16
18
20
Time (h)
Pow
er
(kW
)
House 1 (kW)
House 2 (kW)
House 3 (kW)
0 5 10 15 200
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
Time (h)
Cost
(€)
House 1
House 2
House 3
0 5 10 15 200
0.4
0.8
1.2
1.5
Time (h)
Po
we
r (k
Wh
)
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
102
6.2.1.3 Scenario III - Parameterization for cost benefits
In the third scenario the penalty values of the parameters related with consumption were
increased. The comfort issues are now less important, with the agents mainly concerned with
lower consumptions and in satisfy the power constraint. With this variation, the soft power
constraint was transformed in a hard constraint.
Table 6.3. Scenario III - Penalty values
Due the scarcity of green resource and the obligation to respect the power limits, A3 was the one
that presented lower consumptions, Figure 6.30, and consequently minor costs, but on the other
hand, the indoor temperature was the most penalized with the highest deviation from the chosen
comfort range. With this parameterization, the consumption profile always maintained inside
bounds, for all TCA’s, Figure 6.26, Figure 6.27 and Figure 6.28. Consequently, in Figure 6.29
can be seen that any or negligible red resource was consumed.
Figure 6.23. Scenario III - Disturbance forecasting and indoor temperature A1.
0 5 10 15 200
0.5
1
1.5
2
Pow
er
(kW
)
0 5 10 15 2018
19
20
21
22
23
Time (h)
Tem
pera
ture
(ºC
)
Disturbance Load A1(kW)
Indoor Temp. (ºC) Temp. Max (ºC) Temp. Min (ºC)
Parameter A1 A2 A3
𝛯 50 50 50
100000 100000 30000
ϕ 20 20 20
10 10 10
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
103
Figure 6.24. Scenario III - Disturbance forecasting and indoor temperature A2.
Figure 6.25. Scenario III - Disturbance forecasting and indoor temperature A3.
Figure 6.26. Scenario III – A1 power profile.
0 5 10 15 200
0.5
1
1.5
2
Pow
er
(kW
)
0 5 10 15 2021
21.5
22
22.5
23
23.5
24
Time (h)
Tem
pera
ture
(ºC
)
Disturbance Load A2(kW)
Indoor Temp. (ºC) Temp. Max (ºC) Temp. Min (ºC)
0 5 10 15 200
0.5
1
1.5
2
2.5
3
Pow
er
(kW
)
0 5 10 15 2020
21
22
23
24
25
Time (h)
Tem
pera
ture
(ºC
)
Indoor Temp. (ºC) Temp. Max (ºC) Temp. Min (ºC)
Disturbance Load A3(kW)
0 5 10 15 20-2
-1
0
1
2
3
Pow
er
(kW
)
0 5 10 15 200
0.5
1
1.5
2
2.5
Time (h)
Pow
er
(kW
)
Power Min (kW) Consumption (kW) Power Max (kW)
Total Available (kW) Red Resource (kW) Green Resource (kW)
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
104
Figure 6.27. Scenario III – A2 power profile.
Figure 6.28. Scenario III – A3 power profile.
Figure 6.29. Scenario III - Global consumption characterization.
0 5 10 15 20-2
-1
0
1
2
Pow
er
(kW
)
0 5 10 15 200
0.5
1
1.5
2
Time (h)
Pow
er
(kW
)
Power Min (kW) Consumption (kW) Power Max (kW)
Total Available (kW) Red Resource (kW) Green Resource (kW)
0 5 10 15 20-2
-1
0
1
2
Pow
er
(kW
)
0 5 10 15 200
0.5
1
1.5
2
Time (h)
Pow
er
(kW
)
Red Resource (kW) Green Resource (kW) Total Available (kW)
Power Max (kW) Power Min (kW) Consumption (kW)
0 5 10 15 200
0.5
1
1.5
2
2.5
Time (h)
Po
we
r (k
W)
Total Available (kW)
Total Red (kW)
Total Green (kW)
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
105
Figure 6.30. Scenario III - Power profile.
Figure 6.31. Scenario III - Heating/cooling total cost.
Therefore the energy costs, Figure 6.31, decreased in all agents, namely 8%, 9% and 54%
comparatively with Scenario I, and 9%, 45% and 25% when compared with Scenario II, as
summarized in Table 6.4.
Table 6.4. Scenario III - Cost comparisons
Scenario A1 A2 A3
I 1,27 1,09 1,27
II 1,39 1,52 0,92
III 1,18 1,00 0,69
0 5 10 15 200
2
4
6
8
10
12
14
Time (h)
Po
we
r (k
W)
House 1 (kW)
House 2 (kW)
House 3 (kW)
0 5 10 15 200
0.2
0.4
0.6
0.8
1
1.2
Time (h)
Co
st
(€)
House 1
House 2
House 3
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
106
6.2.1.4 Scenario IV - Parameterization for comfort benefits
The fourth scenario is focused in maintaining the indoor comfort, all agents want to respect the
established temperature gap regardless the required consumption. To accomplish this goal, the
temperature penalty was significantly increased, and the consumption parameters decreased,
Table 6.5.
Table 6.5. Scenario IV - Penalty values
Parameter A1 A2 A3
𝛯 50000 50000 50000
1 1 3
ϕ 0.2 0.2 0.2
0.1 0.1 0.1
Figure 6.32. Scenario IV - Disturbance forecasting and indoor temperature A1.
0 5 10 15 200
0.5
1
1.5
2
Pow
er
(kW
)
0 5 10 15 2018
19
20
21
22
23
Time (h)
Tem
pera
ture
(ºC
)
Indoor Temp. (ºC) Temp. Max (ºC) Temp. Min (ºC)
Disturbance Load A1(kW)
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
107
Figure 6.33. Scenario IV - Disturbance forecasting and indoor temperature A2
Figure 6.34. Scenario IV - Disturbance forecasting and indoor temperature A3
Figure 6.35. Scenario IV – A1 power profile.
0 5 10 15 200
0.5
1
1.5
2
Pow
er
(kW
)
0 5 10 15 2021
21.5
22
22.5
23
23.5
24
Time (h)
Tem
pera
ture
(ºC
)
Temp. Max (ºC) Temp. Min (ºC) Indoor Temp. (ºC)
Disturbance Load A2(kW)
0 5 10 15 200
0.5
1
1.5
2
2.5
3
Pow
er
(kW
)
0 5 10 15 2020
21
22
23
24
25
Time (h)
Tem
pera
ture
(ºC
)
Disturbance Load A3(kW)
Indoor Temp. (ºC) Temp. Max (ºC) Temp. Min (ºC)
0 5 10 15 20-3
-2
-1
0
1
2
3
Pow
er
(kW
)
0 5 10 15 200
0.5
1
1.5
2
2.5
Time (h)
Pow
er
(kW
)
Power Max (kW) Power Min (kW) Consumption (kW)
Total Available (kW) Red Resource (kW) Green Resource (kW)
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
108
Figure 6.36. Scenario IV - A2 power profile.
Figure 6.37. Scenario IV - A3 power profile.
Figure 6.38. Scenario IV - Global consumption characterization.
0 5 10 15 20
-3
-2
-1
0
1
2
3
Pow
er
(kW
)
0 5 10 15 200
0.5
1
1.5
2
2.5
3
3.5
Time (h)
Pow
er
(kW
)
Power Max (kW) Power Min (kW) Consumption (kW)
Red Resource (kW) Green Resource (kW) Total Available (kW)
0 5 10 15 20-2
-1
0
1
2
Pow
er
(kW
)
0 5 10 15 200
0.5
1
1.5
2
Time (h)
Pow
er
(kW
)
Red Resource (kW) Green Resource (kW) Total Available (kW)
Power Max (kW) Consumption (kW) Power Min (kW)
0 5 10 15 200
0.5
1
1.5
2
2.5
3
3.5
4
Time (h)
Pow
er
(kW
)
Total Available (kW)
Total Red (kW)
Total Green (kW)
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
109
As can be seen in Figure 6.32, Figure 6.33 and Figure 6.34 in all the indoor temperatures are
mostly maintained inside the comfort gap.
As consequence, the power constraints, Figure 6.35, Figure 6.36 and Figure 6.37, are violated
and the red resource consumption increased significantly,
Figure 6.38.
Figure 6.39. Scenario IV - Power profile.
Figure 6.40. Scenario IV - Heating/cooling total cost.
Being now the comfort a priority it’s quite clear that the controller tries to accomplished the pre-
defined comfort range leading to a consumption and cost surplus.
As expected, distinct results were obtained in the several proposed scenarios. Figure 6.41
summarize the obtained costs results for the 24hours simulation.
0 5 10 15 200
2
4
6
8
10
12
14
16
18
20
Time (h)
Po
we
r (k
W)
House 1 (kW)
House 2 (kW)
House 3 (kW)
0 5 10 15 200
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
Time (h)
Cost
(€)
House 1
House 2
House 3
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
110
Figure 6.41. Daily heating/cooling total cost.
Comparing the economic differences between scenarios, the third scenario has shown to be the
most economical; the parameters penalizing the consumption were significantly increased,
leading to a situation of less energy usage. However, this decreased consumption led to a higher
indoor temperature deviation from the established boundaries. On the other hand, Scenario IV
was the most expensive. In this scenario the penalty value in the temperature constraint was
increased, showing that the consumers were only concerned in maintaining the temperature
inside the boundaries. Thus, all the necessary resources were consumed, leading to higher costs,
in order to respect the comfort limits. The possibility of obtaining comfort in detriment of the
cost may be important in various situations. For example: To acclimatize rooms with children or
areas in laboratories and/or hospitals. Remark that, despite this comfort preference, the cost
function (in due proportion given by the parameters) also minimizes all the other terms.
The parameterizations from the first and second scenarios have shown to be the most balanced,
with a compromise between comfort and costs.
6.3 Access order with variable hourly sequence
In this chapter it is considered that the sequence access order may vary hourly (Barata et al.,
2013a). The TCA bids are made according to the consumption needs. Each TCA is isolated (no
thermal interactions are considered), has a known fixed 24 hours consumption profile and it is
established a priority level from 1 (low) to 3 (high), that indicates how important that hour is in
terms of consumption. Thus, the hours with high priority levels indicates high consumption and
consequently a higher bid value, Table 6.7.
1 2 3
0.8
1
1.2
1.4
1.6
1.8
2
TCA's
€
Scenario I
Scenario 2
Scenario 3
Scenario 4
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
111
6.3.1 Algorithm II - Implemented sequential scheme with variable
hourly sequence
As mentioned it is considered that the sequence order was already stablished by a previous
auction, but in this chapter, the access sequence order vary hourly. The hourly access sequence
is storage in AO(NS×HP) as exemplified in Table 6.6 for NS=5 and HP=24. Each agent predicts
the consumption and subtracted it to maximum available communicates that information to the
next on the sequence list. Remark that, only the divisions that thermally interact pass to each
other the information about future indoor temperatures.
Table 6.6. Example of hourly access order sequence to green energy
Access
order
Time (h)
1 2 3 4 … 23 24
TCA 1 TCA 3 TCA 5 TCA 1 … TCA 4 TCA 1
TCA 2 TCA 1 TCA 1 TCA 2 … TCA 5 TCA 2
TCA 5 TCA 2 TCA 3 TCA 5 … TCA 3 TCA 3
TCA 4 TCA 4 TCA 2 TCA 4 … TCA 2 TCA 4
TCA 3 TCA 5 TCA 4 TCA 3 … TCA 1 TCA 5
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
112
Algorithm II - DSM-DMPC Implemented sequential scheme with variable hourly sequence
Required for all TCA’s: Thermal disturbances, comfort temperature gap, hourly constraints
parameters and hourly auction bid.
For all TCA’s Wi initialize:
𝐶𝑤𝑙𝑖 fixed consumption within HP (1 × 𝐻𝑃)
BV bid value by hour inside the HP (1 × 𝐻𝑃)
AO access order to green resource is established within the HP (𝑁𝑆 × 𝐻𝑃).
for k=1 to NC
for i=1 to NS(the number of TCA’s)
Get the access order at current instant, AO(k)
Apply to all TCA 𝑢𝑙𝑖(𝑘 − 1: 𝑘 − 1 + 𝐻𝑃)from k-1 instant to obtain 𝑇𝑙
𝑖(𝑘: 𝑘 + 𝐻𝑃)
Communicate temperature predictions to adjacent TCA’s
Update available green resource
��𝑖(𝑘: 𝑘 + 𝐻𝑃) = ��𝑖(𝑘: 𝑘 + 𝐻𝑃) −∑𝐶𝑤𝑙𝑖(𝑘: 𝑘 + 𝐻𝑃)
𝑁𝑑𝑖
𝑙=1
Calculate the optimal control sequence 𝑢𝑙𝑖(1:𝐻𝑃) solving OPi (5.1)-(5.11) with power
constraints given by 𝑈𝑖 and ��𝑖
Update available green resource values for next TCA
��𝑖+1(𝑘: 𝑘 + 𝐻𝑃) = ��𝑖(𝑘: 𝑘 + 𝐻𝑃) −∑𝑢𝑙𝑖(𝑘: 𝑘 + 𝐻𝑃)
𝑁𝑑𝑖
𝑙=1
end for
Update batteries available energy
end for
Remark: generically, 𝑋(𝑘: 𝑘 + 𝐻𝑃) represents a line vector (1 × 𝐻𝑃) containing values from
𝑥(𝑘) to 𝑥(𝑘 + 𝐻𝑃), and 𝑌(𝑝, 𝑘: 𝑘 + 𝐻𝑃), represents the line p of a matrix (𝑃 × 𝐻𝑃) containing
values from y(𝑝, 𝑘) to y(𝑝, 𝑘 + 𝐻𝑃).
6.3.2 Results
Figure 6.42 presents the outdoor temperature and Figure 6.43 to Figure 6.45, presents the
known fixed 24 hours consumption profile and priority level from the three TCA’s.
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
113
Figure 6.42. Outdoor temperature forecasting (Toa).
Figure 6.43. Fixed consumption profile 𝐶𝑤11, and priority level of TCA1.
Figure 6.44. Fixed consumption profile 𝐶𝑤12, and priority level of TCA2.
0 5 10 15 2015
20
25
30
35
40
Time (h)
Te
mp
era
ture
(ºC
)
Temp. Outside
0 5 10 15 200
1
2
3
4
Pow
er
(kW
)
0 5 10 15 201
1.5
2
2.5
3
Time (h)
Level
Fixed consumption A1(kW)
Priority Level
0 5 10 15 200
1
2
3
4
Pow
er
(kW
)
0 5 10 15 201
1.5
2
2.5
3
Time (h)
Level
Fixed consumption A2(kW)
Priority Level
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
114
Figure 6.45. Fixed consumption profile 𝐶𝑤13, and priority level of TCA3.
As mentioned, the priority level presented is stablished according the consumption needs. Thus,
higher consumption represents higher priority levels, and consequently higher bid values, as
presented in Table 6.7.
Table 6.7. Bid value for each consumption level by agent.
Consumption Priority Level House 1 House 2 House 3
0-1 kW 1 2/50.09 3/50.09 1/20.09
1-2 kW 2 7/100.09 4/50.09 2/30.09
>2 kW 3 8.5/100.09 9/100.09 3/40.09
The bid value establishes an order to access to the resource, being the green resource
consumption made by the agents sequentially by that order. The first agent consumes and the
remainder green resource is passed to the next agent as the maximum green available resource.
As mentioned, when the green resource becomes insufficient to satisfy all the demand, the red
is available. The red resource consumption implies a penalty in the final cost function (5.1) due
to the soft constraint violation imposed by the maximum available green resource is exceeded.
A scheme of the system implemented is shown in the next picture.
Available green power
-
TgreenU
1U1
11 uU ….
-
1
1Cw2
1Cw
2
12 uU 2U
Figure 6.46. Implemented system.
0 5 10 15 200
1
2
3
4
5
Pow
er
(kW
)
0 5 10 15 201
1.5
2
2.5
3
Time (h)
Level
Priority Level
Fixed consumption A3(kW)
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
115
𝐶𝑤𝑙𝑖 = [𝑐𝑤𝑙
𝑖(𝑘), … , 𝑐𝑤𝑙𝑖(𝑘 + 𝐻𝑃)]
𝑇 (6.1)
𝑈𝑇𝑔𝑟𝑒𝑒𝑛 = [|𝑢𝑇𝑔𝑟𝑒𝑒𝑛(𝑘)|, … , |𝑢𝑇𝑔𝑟𝑒𝑒𝑛(𝑘 + 𝐻𝑃)|]𝑇 (6.2)
𝑢𝑙𝑖 = [|𝑢𝑙
𝑖(𝑘)|, … , |𝑢𝑙𝑖(𝑘 + 𝐻𝑃)|]
𝑇 (6.3)
where, for a generic TCA i at the control horizon, ��𝑖 represents the green available resource for
indoor comfort , UTgreen represents the green available total resource, 𝐶𝑤𝑙𝑖 the fixed consumption
profile from TCA i division l and 𝑢𝑙𝑖 the used power to heating/cooling the division l inside
TCA i, that results from the optimization program.
It is considered that all houses have the same outdoor temperature presented in Figure 6.42. The
thermal characteristics, load disturbances profile and comfort temperature bounds are different
for all houses. Agents can also have distinct penalties on power and temperature constraints
violations, they can hourly privilege comfort or cost according to consumer choice. Here, is
assumed that the penalty values of each agent are always the same. Table 6.8, shows the used
parameters.
Table 6.8. Scenario parameters.
Parameter A1 A2 A3 Units
Req 50 25 75 ºC/kW
Ceq 9.2103 4.6103 11103 kJ/ºC
𝛯 100 100 300 -
500 200 300 -
ϕ 2 2 2 -
1 1 1 -
t 1 1 1 h
HP 24 24 24 -
T(0) 24 23 24 ºC
As a result from the hourly variation of the priority levels, the bid values also vary hourly
yielding distinguish access orders as showed in Figure 6.47.
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
116
Figure 6.47. Access order.
Figure 6.48. Disturbance forecasting and indoor temperature A1.
Figure 6.49. Disturbance forecasting and indoor temperature A2.
0 5 10 15 201
2
3
Time (h)
Ord
er
A1 A3 A2
0 5 10 15 200
0.5
1
1.5
Time (h)
Pow
er
(kW
)
Disturbance Load A1(kW)
0 5 10 15 2020
22
24
26
Time (h)
Tem
pera
ture
(ºC
)
Indoor Temp.
Temp. Min.
Temp. Max.
0 5 10 15 200
0.5
1
Pow
er
(kW
)
Disturbance Load A2(kW)
0 5 10 15 2020
22
24
26
Time (h)
Tem
pera
ture
(ºC
)
Indoor Temp. Temp. Min. Temp. Max.
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
117
Figure 6.50. Disturbance forecasting and indoor temperature A3.
The thermal disturbances forecasts and the indoor temperature with its constraints for the TCA,
A1, A2 and A3 have the profile presented in, Figure 6.48, Figure 6.49 and Figure 6.50
respectively, and despite it and the access order variability, it can be seen in that the indoor
temperatures are always inside their narrow bounds. Taking advantage of the predictive
knowledge of the disturbance and making use of the space thermal storage, it can also be seen
that the MPC treats the indoor temperature before the disturbance beginning.
The Figure 6.51, Figure 6.52 and Figure 6.53 show for TCA, A1, A2 and A3 respectively, the
used power to acclimatize the spaces and the available green resource to do it. Remark that to
the green resource value must be subtracted the fixed consumption and then, the remainder
represents the maximum available power to provide comfort. Thus, in these figures the “Power
Max” represents the power constraint, ��𝑖 = ��𝑖−1 − 𝑢𝑖−1𝑙 − 𝐶𝑤𝑙
𝑖, “Green resource” ��𝑖−1 −
𝑢𝑖−1𝑙 , and ”Consumption” represents 𝑢𝑖
𝑙.
Figure 6.51. Consumption and constraints to heat/cool A1.
0 5 10 15 200
0.1
0.2
0.3
0.4
0.5
Pow
er
(kW
)
Disturbance Load A3(kW)
0 5 10 15 2020
22
24
26
Time (h)
Tem
pera
ture
(ºC
)
Temp. Min. Temp. Max. Indoor Temp.
0 5 10 15 20-2
-1
0
1
2
3
4
5
Time (h)
Pow
er (
kW)
Green resource (kW) Power Max (kW) Power Min (kW) Consumption (kW)
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
118
Figure 6.52. Consumption and constraints to heat/cool A2.
Figure 6.53. Consumption and constraints to heat/cool A3.
Figure 6.54. Global consumption.
In Figure 6.51 to Figure 6.53, it can be seen that the comfort constraints are respected, the used
power to heat/cool the space is maintained inside the constrained bounds. Note that when the
0 5 10 15 20-4
-3
-2
-1
0
1
2
3
4
5
Time (h)
Po
wer
(k
W)
Power Min (kW) Power Max (kW) Consumptiom (kW) Green resource (kW)
0 5 10 15 20-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
3
3.5
Time (h)
Po
wer
(k
W)
PowerMax (kW) Green resource (kW) Consumptiom (kW) Power Min (kW)
0 5 10 15 200
1
2
3
4
5
6
7
8
Time (h)
Pow
er
(kW
)
Red Resouce (kW) Green Resource (kW) Total Available (kW)
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
119
resource is null the consumption is also null. Figure 6.54 shows the global used power and
characterized it in terms of type of consumed power.
𝑈𝑢𝑠𝑒𝑑(𝑘) =∑[|𝑢𝑙𝑖(𝑘)| + 𝐶𝑤𝑙
𝑖].
It can be seen in Figure 6.54, that in the most demanding periods, the maximum green available
resource is exceeded and the red resource must be consumed.
Figure 6.55. Heating/cooling total cost.
Figure 6.55 illustrates the effectiveness of the approach and demonstrates the advantage of the
auction. For each one of the agents it can be seen that the “Real Cost” is much lower than the
cost of not to bid in auction and only consume the red resource “Red Cost”.
6.4 Conclusions
In this chapter, a distributed MPC control technique was presented in order to provide thermal
house comfort. The solution obtained solves the problem of controlling multiple subsystems
dynamically coupled and also exposed to a coupled constraint. Each subsystem solves its own
problem by involving its own state predictions or from neighbourhoods and the shared
constraints. It could be observed through the simulations and result’s analysis that suitable
dynamic performances were obtained. The built system is flexible in two ways. It allows in each
auction the agents to bid higher or lower in accordance with their needs and, by changing the
penalty values during the day, consumers can shift hourly between indoor comfort and lower
costs. Knowing in advance the disturbance forecasts allows agents to make their bid and
achieve significant savings at the end of the day.
The distributed MPC control technique, along with a thermal-electrical modular scheme, was
validated in order to provide thermal house comfort, with strong presence of intermittent/limited
0 5 10 15 200
0.2
0.4
0.6
0.8
1
1.2
1.4
Time (h)
Cost
(€)
Red Cost A1
Red Cost A2
Red Cost A3
Real Cost A1
Real cost A2
Real Cost A3
DMPC FOR THERMAL HOUSE COMFORT WITH SEQUENTIAL ACCESS AUCTION
120
RES environments. The approach spots a control problem of multiple subsystems dynamically
coupled and exposed to coupled constraint. Each subsystem solves its own problem by
involving its own state predictions and the state predictions of adjacent rooms, available with
communication interchange and shared constraints. The changing in penalty values allows
consumer to pick between indoor comfort and lower costs.
121
Chapter 7
7 DMPC for Thermal House Comfort
with Sliding Load
7.1 Introduction
The desired approach here presented intends to take advantage from the innovative technology
characteristics provided by future SGs (Korolija et al., 2011). In the smart world, simple
household appliances, like dishwashers, clothes dryers, heaters or air conditioners are assumed
to be fully controllable in order to achieve the network maximum efficiency. Renewable
energies will be a common presence and any kWh provided by these technologies should not be
wasted. Active Demand-Side Management will control the loads in order to adapt them to the
available renewable energy source. As mentioned in Chapter 2.2, DSM studies are focused in
the development of load control manipulation models (EIA, 2014; Favre-Perrod et al., 2009)
and electricity incentive prices to promote load management (Kosek et al., 2013; Chen, 2010).
In buildings, DSM is based on an effective reduction of the energy needs by changing the shape
and amplitude consumer’s load diagrams. So, DSM can involve a combination of several
strategies; pricing, load management curves and energy conservation are implemented for a
more energy efficient use.
Load shifting is considered a common practice in the management of electricity supply and
demand, where the peak energy use is shifted to less busy periods. Properly done, load shifting
helps meeting the goals of improving energy efficiency and reducing emissions, smoothing the
daily peaks and valleys of energy use and optimising existing generation assets. With new
technological advances, DR programmes may shift loads by controlling the function of air
CONTROL FOR THERMAL HOUSE COMFORT WITH SLIDING LOAD
122
conditioners, refrigerators, water heaters, heat pumps and other similar electric loads at
maximum demand times. The work here presented is distinct because provides an integrative
solution which is able to, in a distributed network with multiple TCAs, adjust the demand to an
intermittent limited energy source, using load shift and maintaining the indoor comfort, (Barata
et al., 2013c) and more detailed in (Barata et al., 2014c). Figure 7.1, exemplifies a shifting load
communication infrastructure.
Figure 7.1. Shifting load communication infrastructure.
Thus, due the specificities made in this chapter, the following assumptions are added to the ones
made in Chapter 6.1:
Each division selects the load value (LV), the duration (LVd), the turned on time (ToT)
and the “sliding level” (SL) of the “shifted loads”. The SL indicates that the load can
be turned on SL hours before and after the chosen ToT;
The next picture illustrates the shifting load characteristics.
Figure 7.2. Implemented shifting load scheme.
Therefore, using the demand side management, by negotiating the energy price and the
consumer comfort, the distributed loads can be managed and shifted in order to guarantee the
Shifting load
Controler
Shiftable loadsHVAC, Dishwaher, PHEV,…
Non-Shiftable loadsLigthing, Frigde ,…
Communication Line
Power Line
Consumer
Parameterization
ToT
LVd
LV
SL=1 SL=1SL=2 SL=2
1
...
2
...
LV
...
Po
wer
(k
W)
Time (h)
CONTROL FOR THERMAL HOUSE COMFORT WITH SLIDING LOAD
123
satisfaction of all constraints. The global system includes an auction (provided by the MO) that,
in accordance with the bid value made by the houses/agents, defines an order to access the
green resource. The green resource consumption is made by the agents sequentially by the
auction order and the information about the remaining green resource is passed to the next agent
as the maximum green available resource. As mentioned, when green resource becomes
insufficient to satisfy all the demand, the red is available.
7.2 Algorithm III – Implemented shifting and loads allocation
scheme
Each one of the systems starts by choosing their loads characteristics, LV, LVd, ToT and SL. With
this data, all the possible loads schedule combinations (PLSCS) are establish (see Figure 7.7
e.g.). At each time step, it’s verified if inside the predictive horizon, any PLSCS exceeds the
maximum available green energy iU max . The sequences that are at any instant above the
iU max limit are removed, and the remaining are the feasible load schedule combinations
(FLSCS). The FLSCS are subtracted to iLU , and the resulting in a set of combinations iU that
are tested in the minimization problem as maximum available green resource for comfort (5.1).
The hypothesis that provided less consumption is chosen. Once one sequence is started, all the
others that are different until the current step time are eliminated until the final load sequence is
chosen, FLSeq. The total consumption by division and house at any instant can be written as
(7.1) and (7.2) respectively, and is pictured in Figure 7.4.
Figure 7.3. Total consumption characterization.
Remark that the total available for comfort for each house (7.3) is used as constraint in the
optimization problem. The equations assume the following form,
𝑃𝑙𝑖(𝑘) = |𝑢𝑙
𝑖(𝑘)| + 𝐶𝑤𝑙𝑖(𝑘) + 𝐹𝐿𝑆𝑒𝑞𝑙
𝑖(𝑘), (7.1)
𝑃𝑖(𝑘) =∑𝑃𝑙𝑖(𝑘)
𝑁𝑑𝑖
𝑙=1
, (7.2)
CONTROL FOR THERMAL HOUSE COMFORT WITH SLIDING LOAD
124
𝑈𝑖(𝑘) = 𝑢 𝑇𝑔𝑟𝑒𝑒𝑛(𝑘) −∑(𝐶𝑤𝑙𝑖(𝑘) − 𝐹𝐿𝑆𝑒𝑞𝑙
𝑖(𝑘))
𝑁𝑑𝑖
𝑙=1
−∑𝑃𝑗(𝑘)
𝑁𝑆
𝑗=1𝑖≠𝑗
. (7.3)
A simplified scheme of the implemented optimization problem is shown in the next picture,
Figure 7.4, followed by the implemented DSM-DMPC algorithm.
TgreenUU 1max
--
iU
OPi
iLU
iTgreeni PUU 1max
--
--
OPNS
1
1
maxS
S
N
i
iTgreenN PUU
Wi
SNW
iPSLC
iNd
l
i
lCw
1iU
1iLU
12max iiTgreeni PPUU
Wi+1
1iPSLC
1
1iNd
l
i
lCw
OPi+1
SNU
SNLU1iPSLC
SN
S
Nd
l
N
lCw
Figure 7.4. Implemented power distribution scheme starting in the Optimization Problem 1
(OPi) to OPNS.
CONTROL FOR THERMAL HOUSE COMFORT WITH SLIDING LOAD
125
Algorithm III - DSM-DMPC pseudocode prototype algorithm
For all houses Wi initialize:
LV load value
LVd the duration
ToT the turned on time
SL sliding level
iPSLC possible schedule loads combinations (𝑛𝑃𝐶 ×𝐻𝑃) with nPC the number of possible
combinations
𝐶𝑤𝑙𝑖 fixed consumption within HP (1 × 𝐻𝑃)
BV bid value by hour inside the HP (1 × 𝐻𝑃)
AO access order to green resource is established within the HP (𝑁𝑆 × 𝐻𝑃).
for k=1 to NC
for i=1 to NS(the number of TCA’s)
Get the access order at current instant, AO(k)
Apply to all TCA 𝑢𝑙𝑖(𝑘 − 1: 𝑘 − 1 + 𝐻𝑃)from k-1 instant to obtain 𝑇𝑙
𝑖(𝑘: 𝑘 + 𝐻𝑃)
Communicate temperature predictions to adjacent TCA’s
Update available green resource
��𝐿𝑖(𝑘: 𝑘 + 𝐻𝑃) ← ��𝑚𝑎𝑥𝑖(𝑘: 𝑘 + 𝐻𝑃) −∑𝐶𝑤𝑙𝑖(𝑘: 𝑘 + 𝐻𝑃)
𝑁𝑑𝑖
𝑙=1
Built table with all FSLCS
𝐹𝐿𝑆𝐶𝑖(𝑛𝐹𝐶 , 𝑘: 𝑘 + 𝐻𝑃) ← 𝑃𝐿𝑆𝐶𝑖(𝑛𝐹𝐶 , 𝑘: 𝑘 + 𝐻𝑃) − ��𝐿𝑖(𝑘: 𝑘 + 𝐻𝑃) ≥ 0
for t=1 to nFC (number of feasible combinations)
Calculate the optimal control sequence ui(1:HP) solving OPi (5.1)-(5.11) with power
constraint given by ��𝑖(𝑘: 𝑘 + 𝐻𝑃) ← 𝐹𝐿𝑆𝐶𝑖(𝑡, 𝑘: 𝑘 + 𝐻𝑃)
𝑈𝑝𝑟𝑒𝑑𝑖(𝑡) ← ∑𝑢𝑖(1:𝐻𝑃)
if 𝑈𝑝𝑟𝑒𝑑𝑖(𝑡) < 𝑈𝑝𝑟𝑒𝑑𝑖(𝑡 − 1) then
𝐹𝐿𝑆𝑒𝑞𝑖(1: 𝑘) ← 𝐹𝑆𝐿𝐶𝑖(𝑡, 𝑘: 𝑘 + 𝐻𝑃)
end if
end for
Eliminate from 𝑃𝐿𝑆𝐶𝑖 all the sequences that are different from 𝐹𝐿𝑆𝑒𝑞𝑖(1: 𝑘)
Update available green resource values for next TCA
��𝑖+1(𝑘: 𝑘 + 𝐻𝑃) = ��𝑖(𝑘: 𝑘 + 𝐻𝑃) −∑𝑢𝑙𝑖(𝑘: 𝑘 + 𝐻𝑃)
𝑁𝑑𝑖
𝑙=1
end for
Update batteries available energy
end for
CONTROL FOR THERMAL HOUSE COMFORT WITH SLIDING LOAD
126
As mentioned the algorithm is sequential, and for a better understanding of the implemented
power distribution scheme, the access order is W1, W2 and so on. Therefore, for a certain instant,
is considered that W1 was the one that made the highest bid, W2 made the second highest, always
sequentially until 𝑊𝑁𝑆 . In Figure 7.4, the available power for W1 is given by the predicted green
total available resource 𝑈𝑚𝑎𝑥1 = 𝑈𝑇𝑔𝑟𝑒𝑒𝑛 at the control horizon (7.5), and them the fixed
consumption (7.4) is subtracted resulting in 𝑈𝐿𝑖. As mentioned, the PLSCS are compared inside
the predictive horizon with 𝑈𝐿𝑖, and the ones that exceeds at any instant limit are removed,
and the remaining are FLSCS. The FLSCS are subtracted to , and the resulting in a set of
combinations 𝑈𝑖.. These combinations are the power constraint (5.10) and are tested in the OPi
(5.1). The combination that generate lower consumption, ui , is chosen. Then, the information
about the available green energy is passed for the next house, 𝑈𝑚𝑎𝑥𝑖.
For a generic Agent i at the control horizon 𝑈𝑇𝑔𝑟𝑒𝑒𝑛 represent the green available total resource,
𝐶𝑤𝑙𝑖 the fixed consumption profile and 𝑢𝑙
𝑖the used power to heating/cooling the space that
results from the optimization program. These parameters are express by the vectors (7.4)-(7.6),
𝐶𝑤𝑙𝑖 = [𝑐𝑤𝑙
𝑖(𝑘), . . . , 𝑐𝑤𝑙𝑖(𝑘 + 𝐻𝑃)]
𝑇, (7.4)
𝑈𝑇𝑔𝑟𝑒𝑒𝑛 = [|𝑢𝑇𝑔𝑟𝑒𝑒𝑛(𝑘)|, . . . , |𝑢𝑇𝑔𝑟𝑒𝑒𝑛(𝑘 + 𝐻𝑃)|]𝑇, (7.5)
𝑢𝑙𝑖 = [|𝑢𝑙
𝑖(𝑘)|, . . . , |𝑢𝑙𝑖(𝑘 + 𝐻𝑃)|]
𝑇. (7.6)
In the built algorithm it is considered that the access order to the green resource is established
hourly according to bid value made in auction by each agent, and therefore, at each instant the
defined access sequence must be applied. Another feature provided by the implemented system
is that each house can have different hourly penalties, allowing the consumer to choose between
more/less comfort and cost during the day.
7.3 Results
The presented results were obtained with an optimization MATLAB® routine that finds a
constrained minimum of a quadratic cost function that penalizes the sum of the several
objectives (5.1).
CONTROL FOR THERMAL HOUSE COMFORT WITH SLIDING LOAD
127
7.3.1 One house scenario
To simplify better understand the used approach, the first results here present show only the
shifting loads procedure for one house represented by one division with thermal disturbance
(QPd), Figure 7.6 (no fixed consumption profile and storage are considered). Table 7.1 shows the
used scenario parameters. The outdoor temperature forecast, Figure 7.5, considers 90%
accuracy to within +/- 2°C on the next day.
Figure 7.5.Outdoor temperature forecasting (Toa).
Table 7.1.Scenario parameters
Req(ºC/kW) Ceq(kJ/ºC) Ξ Ψ ϕ 𝜑 t(h) HP NC T(0) (ºC)
50 9.2103 500 500 2 1 1 24 24 21
Figure 7.6. Thermal disturbance forecasting.
2 4 6 8 10 12 14 16 18 20 22 2415
20
25
30
35
40
Time (h)
Te
mp
era
ture
(ºC
)
Outdoor Temp.
2 4 6 8 10 12 14 16 18 20 22 240
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
Time (h)
Pow
er
(kW
)
Disturbance thermal load (kW)
CONTROL FOR THERMAL HOUSE COMFORT WITH SLIDING LOAD
128
The Loads that can be daily shifted have the characteristics present in the next Table 7.2, and
Figure 7.7 shows all the possible 56 loads combinations in the 24hours period.
Table 7.2. Shifted loads characteristics
Loads LV (kW) LVd (h) ToT (h) SL (h)
Load 1 3 2 8 3
Load 2 2 3 18 4
Figure 7.7. Possible loads schedule combinations (PLSCs).
The system tests all combinations present in Figure 7.7, and as mentioned above, the hypotheses
that do not respect the maximum predicted green resource are initially discharge, and the
remaining ones the FLSCS, Table 7.3, are tested in (5.1).
The sequence FLSeq, where mostly green energy is consumed, the costs are lower and the
indoor comfort range is respected is found.
510
1520
10
20
30
40
50
0
1
2
3
Time (h)
PLSCS
Pow
er
(kW
)
CONTROL FOR THERMAL HOUSE COMFORT WITH SLIDING LOAD
129
Table 7.3. Feasible Loads Sequence Combinations
FL
SC
Time (h)
1 2 3 4 5 6 7 8 9 10 11 12
1 0 0 0 0 0 0 0 0 3 3 3 0
2 0 0 0 0 0 0 0 0 3 3 3 0
3 0 0 0 0 0 0 0 0 3 3 3 0
4 0 0 0 0 0 0 0 0 3 3 3 0
FL
SC
Time (h)
13 14 15 16 17 18 19 20 21 22 23 24
1 0 2 2 2 2 0 0 0 0 0 0 0
2 0 0 2 2 2 2 0 0 0 0 0 0
3 0 0 0 2 2 2 2 0 0 0 0 0
4 0 0 0 0 2 2 2 2 0 0 0 0
In Figure 7.8 are the total energy costs of the FLSCS shown in Table 7.3, and can be seen that
the chosen sequence is the less expensive.
Figure 7.8. Total energy costs of FLSCS.
Figure 7.9, show the chosen load sequence, FLSeq, and the available green energy to allocate
the loads.
1 2 3 40.265
0.266
0.267
0.268
0.269
0.27
0.271
0.272
0.273
0.274
Hypothesis
€
Total Cost Hypothesis (€)
CONTROL FOR THERMAL HOUSE COMFORT WITH SLIDING LOAD
130
Figure 7.9. Maximum available green energy and chosen sequence.
In order to minimize the energy costs by consuming only green resource, the implemented
algorithm chooses the gaps that fit properly in the maximum available green energy.
Figure 7.10. Indoor temperature.
The comfort limits varies during the 24h period, and Figure 7.10 shows that the indoor
temperature is always maintained inside the comfort limits being the optimization problem able
to respect the temperature and power constraint Figure 7.11.
The periods between 9-11h and 15-18h are extremely demanding, all green energy is consumed
by the shifted loads, with no remaining one for comfort proposes.
Although, Figure 7.10 shows that in that periods the algorithm choose to not use the red
resource and, taking advantage of the prediction horizon, pre-heat or pre-cool the spaces when
only renewable resource is available.
2 4 6 8 10 12 14 16 18 20 22 240
0.5
1
1.5
2
2.5
3
Time (h)
Pow
er
(kW
)
Initial Power Max.
Load Sequence
2 4 6 8 10 12 14 16 18 20 22 2421
21.5
22
22.5
23
23.5
24
Time (h)
Tem
pera
ture
(ºC
)
Indoor Temp.
Temp. Max.
Temp. Min.
CONTROL FOR THERMAL HOUSE COMFORT WITH SLIDING LOAD
131
Figure 7.11. Used power to heat/cool the space and the maximum green resource available for
comfort.
7.3.2 Distributed scenario
It is considered that all houses are represented by one division and have the same outdoor
temperature presented in Figure 7.5. The used parameters are presented in Table 7.4.
Table 7.4. Distributed scenario parameters
Parameter A1 A2 A3 Units
Req 50 25 75 ºC/kW
Ceq 9.2103 4.6103 11103 kJ/ºC
𝛯 100 100 300 -
𝝍 500 200 300 -
𝝓 2 2 2 -
𝝋 1 1 1 -
t 1 1 1 h
HP 24 24 24 -
T(0) 21 23 24 ºC
2 4 6 8 10 12 14 16 18 20 22 24-3
-2
-1
0
1
2
3
Time (h)
Pow
er
(kW
)
Consumption (kW) Power Max (kW) Power Min (kW)
CONTROL FOR THERMAL HOUSE COMFORT WITH SLIDING LOAD
132
The batteries capacity is 3kWh. To incentive the clean resource consumption, it is considered
that the green energy price per kWh has a maximum auction value (0.09€/kWh) always cheaper
than the red energy price (0.17€/kWh). Table 7.5 shows the bid value for each one of the
priority levels that, as mentioned, are established according the fixed consumption profile.
Table 7.5. Bid value for each consumption level by TCA
Consumption Priority Level TCA 1 TCA 2 TCA 3
0-1 kW 1 2/50.09 3/50.09 1/20.09
1-2 kW 2 7/100.09 4/50.09 2/30.09
>2 kW 3 8.5/100.09 9/100.09 3/40.09
The load disturbances profile, comfort temperature bounds and shifted loads characteristics are
different for all houses, Table 7.6.
Table 7.6. Shifted loads characteristics for distributed scenario
TCA Loads LV (kW) LVd (h) ToT (h) SL (h)
1 Load 1 1 2 7 1
Load 2 2 4 18 2
2 Load 1 2 3 9 1
Load 2 2 2 21 2
3 Load 1 3 3 8 1
Load 2 3 3 13 1
In all houses, the fixed consumption profile, 𝐶𝑤11, 𝐶𝑤1
2 and 𝐶𝑤13 and its priority level, are
known within a 24h period and are depicted in Figure 7.12, Figure 7.13 and Figure 7.14.
Figure 7.12. Fixed consumption profile 𝐶𝑤11, and priority level of TCA1.
1 3 5 7 9 11 13 15 17 19 21 23 240
0.5
1
1.5
2
2.5
3
Pow
er (
kW)
1 3 5 7 9 11 13 15 17 19 21 23 241
1.5
2
2.5
3
Time (h)
Leve
l
Fixed consumption A1(kW)
Priority Level
CONTROL FOR THERMAL HOUSE COMFORT WITH SLIDING LOAD
133
Figure 7.13. Fixed consumption profile 𝐶𝑤12, and priority level of TCA2.
Figure 7.14. Fixed consumption profile 𝐶𝑤13, and priority level of TCA3.
Figure 7.15. Access order.
1 3 5 7 9 11 13 15 17 19 21 23 240
1
2
3
4
Pow
er
(kW
)
1 3 5 7 9 11 13 15 17 19 21 23 241
1.5
2
2.5
3
Time (h)
Leve
l
Priority Level
Fixed consumption A2(kW)
1 3 5 7 9 11 13 15 17 19 21 23 240
1
2
3
4
Pow
er
(kW
)
Fixed consumption A3(kW)
0 5 10 15 201
1.5
2
2.5
3
Time (h)
Leve
l
Priority Level
2 4 6 8 10 12 14 16 18 20 22 241
2
3
Time (h)
Ord
er
A3 A2 A1
CONTROL FOR THERMAL HOUSE COMFORT WITH SLIDING LOAD
134
As a result from the hourly variation of the priority levels, the bid values also vary hourly
yielding distinguish access orders as showed in Figure 7.15.
The thermal disturbance profile is known within a 24 hour period, and is related with thermal
loads generated by occupants, direct sunlight, electrical devices or doors and windows aperture
to recycle the indoor air.
Figure 7.16. Disturbance forecasting and indoor temperature A1.
Figure 7.17. Disturbance forecasting and indoor temperature A2.
1 3 5 7 9 11 13 15 17 19 21 23 240
0.2
0.4
0.6
0.8
1
Pow
er
(kW
)
1 3 5 7 9 11 13 15 17 19 21 23 2418
20
22
24
26
Tem
pera
ture
(ºC
)
Time (h)
Indoor Temp. Temp. Max. Temp. Min.
Disturbance Load A1(kW)
1 3 5 7 9 11 13 15 17 19 21 23 240
0.2
0.4
0.6
0.8
1
Pow
er
(kW
)
Time (h)
Disturbance Load A2(kW)
1 3 5 7 9 11 13 15 17 19 21 23 2420
21
22
23
24
25
26
Tem
pera
ture
(ºC
)
Temp. Max. Temp. Min. Indoor Temp.
CONTROL FOR THERMAL HOUSE COMFORT WITH SLIDING LOAD
135
Figure 7.18. Disturbance forecasting and indoor temperature A3.
In Figure 7.16, Figure 7.17 and Figure 7.18, it can be seen that the comfort constraints are
respected, the indoor temperature is always inside the comfort zone in all houses. Taking
advantage of the predictive knowledge of the thermal disturbance and making use of the space
thermal storage, it can also be seen that in all houses the MPC treats the indoor temperature
before the thermal disturbance beginning. In Figure 7.19 it can be seen that the shifted loads
were located in zones with mostly green energy available. Note that when the used power is
above the daily maximum green available resource, means that the red resource was consumed.
The used power to heat/cool the space is maintained inside the constrained bounds and when the
green energy is null the used power is also null, Figure 7.20, Figure 7.22 and Figure 7.24 for
TCA1, TCA2 and TCA3 respectively.
Remark that the fixed consumption, 𝐶𝑤11, 𝐶𝑤1
2 and 𝐶𝑤13 represent the base in the power profile
in Figure 7.19, Figure 7.21 and Figure 7.23.
Figure 7.19. Power profile A1.
0 5 10 15 200
0.1
0.2
0.3
0.4
0.5
Pow
er
(kW
)
Disturbance Load A3(kW)
0 5 10 15 2021
22
23
24
25
26
Time (h)
Tem
pera
ture
(ºC
)
Temp. Max. Temp. Min. Indoor Temp.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 240
1
2
3
4
5
6
7
8
9
10
Time (h)
Pow
er
(kW
)
A1
Red consumption (kW)
Control Input (kW)
Load Sequence (kW)
Fixed consumption forecast (kW)
Power Constraint (Max Available) (kW)
CONTROL FOR THERMAL HOUSE COMFORT WITH SLIDING LOAD
136
Remark that, for example, at time instant t=7, three different types of energy utilization are
used. The base, in dark grey, is fulfil with the fixed consumption, above is the shifted load and
on top is the used power for comfort. In this instant the total consumption is maintained within
the power constraint. On the other hand, at instant t=9, with no available green power, all the
fixed consumption of 2kW is made with red resource.
Figure 7.20. Control input profile A1
In Figure 7.21, it can be seen that the shifted loads of house 2 were located in zones with mostly
green energy available.
Figure 7.21. Power profile A2.
The used power to heat/cool the space is always maintained inside the constrained bounds been
the clean resource consumed only when is available. Note that when the used power to satisfy
all the demand is above the daily maximum green available resource, means that the red
resource was consumed, Figure 7.21.
1 3 5 7 9 11 13 15 17 19 21 23 24-8
-6
-4
-2
0
2
4
6
8
Time (h)
Pow
er
(kW
)
Power Min (kW) Power Max (kW) Consumption (kW)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 240
1
2
3
4
5
6
7
8
9
Time (h)
Po
we
r (k
W)
A2
Red consumption (kW)
Control Input (kW)
Load Sequence
Fixed consumption forecast (kW)
Power Constraint (Max Available) (kW)
CONTROL FOR THERMAL HOUSE COMFORT WITH SLIDING LOAD
137
Figure 7.22. Control input profile A2.
Figure 7.23 shows that the chosen FLSeq3 is located here the consumption of red resource is
obliged. Due the access order imposed by the auction, the maximum available energy may
change hourly, and by this fact the available resource prediction is not as effective as with one
house only. Also, the SL=1 of both loads of house 3 Table 7.6, is obvious a conditionality and
an extra restriction on our optimization algorithm.
The power constrained bounds are respected been the consumed made only when the clean
resource is available, Figure 7.24.
Figure 7.23. Power profile A3.
2 4 6 8 10 12 14 16 18 20 22 24-8
-6
-4
-2
0
2
4
6
8
Time (h)
Pow
er
(kW
)
Consumption (kW) Power Max (kW) Power Min (kW)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 240
1
2
3
4
5
6
7
8
9
10
Time (h)
Pow
er
(kW
)
A3
Red consumption (kW)
Control Input (kW)
Load Sequence
Fixed consumption forecast (kW)
Power Constraint (Max Available) (kW)
CONTROL FOR THERMAL HOUSE COMFORT WITH SLIDING LOAD
138
Figure 7.24. Control input profile A3.
Figure 7.25. Batteries profile.
.
Figure 7.26. Consumption costs.
2 4 6 8 10 12 14 16 18 20 22 24
-6
-4
-2
0
2
4
6
Time (h)
Pow
er
(kW
)
Power Max (kW) Power Min (kW) Consumption (kW)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 240
0.5
1
1.5
2
2.5
3
Time (h)
Pow
er
(kW
)
Batteries Available Power
2 4 6 8 10 12 14 16 18 20 22 240
0.2
0.4
0.6
0.8
1
1.2
1.4
Time (h)
Co
st
(€)
Red Cost A3 Red Cost A1 Red Cost A2 Real Cost A1 Real Cost A3 Real cost A2
CONTROL FOR THERMAL HOUSE COMFORT WITH SLIDING LOAD
139
The batteries profile during the 24h period is show in Figure 7.25. It can be seen that in the most
demanding periods, the energy available in the batteries provide a useful energy support. Figure
7.26 demonstrates the advantage of the auction. For each one of the houses it can be seen that
the “Real Cost” is much lower than the cost of not to bid in auction and only consume the red
resource “Red Cost” at a higher fixed price.
7.4 Conclusions
In this chapter, a distributed MPC control integrative solution was validated in order to provide
thermal house comfort in an environment with strong presence of intermittent/limited RES. The
approach spots a control problem of multiple subsystems (multi-agent) subjected to coupled
constraint solved as a sequence of QP optimization problems for each instant in time.
The approach shows that distributed predictive control is able to provide house comfort within a
DSM policy, based in a price auction and the rescheduling of appliance loads. It is a valid
methodology to achieve reduction in consumption and price. The approach is more effective
with wider periods where the loads are allowed to slide and consequently allocate the most
favourable zone.
CONTROL FOR THERMAL HOUSE COMFORT WITH SLIDING LOAD
140
141
Chapter 8
8 Conclusions and Future Work
Directions
8.1 Conclusion and Summary of Achievements
It is widely accepted that in a nearby future Smart Grids will be a part of our daily life’s. Smart
grids will most likely include decentralized generation, active network generation management,
energy storage and demand management resources, where the actions of all agents connected to
the electricity system can be intelligently integrated aiming for a sustainable, efficient and
secure energy supply system. Widespread communications and control technologies required to
use DR help maintain the supply-demand balance in electricity systems. Smart household
appliances with controllable loads will soon be a common presence in our homes. This active
DSM is able to manage the loads to obtain harmony in demand supply ratio, they also include
load control manipulation models, pricing, with distinct electricity tariffs along the day,
encouraging load management and other approaches which promote energy efficiency and
conservation. Many energy consumers are already able to tie the electricity pricing into their
energy management system, resulting in greater shifts of energy usage. This needs to be
convoyed by the application of intelligent appliances that would facilitate the implementation of
DSM.
Therefore, this work intends to be a solution which is able to respond to all this requirements.
The work has considered environmental issues and pre-establishes variables as a decision-maker
factor. In particular, the work presented a novel multi-agent model-based predictive control
method to manage distributed energy resources (DER) systems from the demand side, when in
CONCLUSIONS AND FUTURE WORK DIRECTIONS
142
presence of limited energy sources with fluctuating output and with energy storage in house-
hold or car batteries. Specifically, the work has presented a solution for thermal comfort which
is able to manage a limited shared energy resource via a demand side management perspective,
using an integrated approach also involving a power price auction and an appliance load
allocation scheme. The control technique has shown to be capable of controlling loads based on
the available supply at a certain time, without significantly compromising the user satisfaction.
The developed building’s energy management system and smart thermostat feature a lot of
programmable settings that allow users and their utilities to maximize energy savings. It is
based on a consumer's schedule or peak time’s energy demand for the utility; and in a
thermostat that pre-programs itself based on weather forecasts to optimize both consumer's
energy savings and grid performance. The system uses, weather conditions forecasts and the
MPC features to anticipate major changes in weather and manages heating and cooling needs in
advance, in the most energy-efficient way.
The term Thermal Control Area (TCA) was presented and described as an entity that is
embedded in a distributed environment, where several others TCA’s are working individually to
achieve their own goal and sharing their information to accomplish a global objective.
An electro-thermal modular scheme was developed to allow an easy building’s modelling with
different plans where adjacent areas with distinct construction features may thermally interact.
The global scenario described in Chapter 5 present all the basic assumptions in which the work
was developed. Also in this chapter the DMPC optimization problem and the Algorithm I which
define the implemented basic sequential scheme, were detailed, formalized and exemplified.
The obtained results showed its suitability.
In Chapter 6 was presented a solution that solves the problem of control of multiple subsystems
dynamically coupled and also stringed to a coupled constraint. Each subsystem solves its own
problem by involving its own state predictions, or from neighbourhoods, and the shared
constraints. It can be observed throughout the simulations and results analysis that suitable
dynamic performances were obtained. The system built in Algorithm II is flexible in two ways.
It allows in each auction the agents bid higher or lower in accordance with their needs and, by
changing the penalty values during the day, the consumer is able to hourly shift between indoor
comfort and lower costs. By knowing in advance the disturbance forecasts, the TCA’s are able
to make their bid and achieve significant savings at the end of the day.
In Chapter 7, a distributed MPC control integrative solution has been validated in order to
provide thermal house comfort in an environment with strong presence of intermittent/limited
RES. The developed Algorithm III has shown that distributed predictive control combined with
a DSM policy is a tool able to provide thermal comfort, based in a price auction and the
CONCLUSIONS AND FUTURE WORK DIRECTIONS
143
rescheduling of appliance loads. It is a valid methodology to achieve less energy consumption
and price reductions. The approach has shown to be more effective with wider periods where
the loads are allowed to slide and consequently allocate the most favourable zone.
The followed table summarize the implemented features which characterises the developed
algorithms.
Table 8.1. Developed algorithm characteristics
Features Algorithm I Algorithm II Algorithm III
Access to green energy Fixed Variable (hourly) Variable (hourly)
Distributed nature Yes Yes Yes
Thermal coupling Yes Yes Yes
Information interchange
between TCA’s Yes Yes Yes
Thermal disturbances Yes Yes Yes
Hourly penalties
variability No Yes Yes
Battery storage No Yes Yes
Fixed load allocation No Yes Yes
Priority consumption
level No Yes Yes
Sliding load allocation No No Yes
8.2 Recommendations for Future Work Directions
In order to efficiently allocate resources to facilitate balancing of both electricity and heat it
would be critical to establish some form of RTP arrangement so users could be fully informed
about the value of heat and electricity at each point in time (and location). This would also
encourage development of appropriate demand side solutions. This will require introduction of
much more sophisticated energy metering (e.g. half hourly metering) and trading functions and
would lead to deployment of information and communication systems to ease control of
generators and loads. The philosophy of liberal markets assumes economic optimums can only
be achieved through negotiations in a free market environment. Because energy is a complex
CONCLUSIONS AND FUTURE WORK DIRECTIONS
144
commodity, a sophisticated commercial structure is necessary to make possible the trading
between each generator or consumer and purchasers of energy and adjuvant services.
A market with hundreds of thousands or even millions of active participants is clearly a major
challenge. All participants will need to access various markets to set up long and short term
energy transactions, secure access to the network and provide ancillary services to operate the
system in a satisfactory manner (if this is to be done in heat energy, heat networks which would
facilitate operation of markets for heat would be required). In order to smooth trading of energy
among a very large number of distributed generators and loads, an electronic energy market
system, supported by the internet, would be needed to be developed.
RES forecasts of long-term power variation should be considered for a future task. As a rational
assumption, it is presumed that RES forecasts are within a short error band, negligible for
predicting horizons of 24 hours, but should be estimated for the effect of power variation during
longer periods.
Typically, distributed RES fluctuations are counter-balanced by the use of battery storage
systems. However, battery systems not only increase the size and cost of RES power systems
but also power fluctuations have the effect of reducing the useful life-time of the battery storage
system. The use of the battery storage system can, therefore, be fully optimised, minimising
storage requirements, avoiding damage to the battery bank and reducing power fluctuations with
the consequent useful life of the battery extent . These issues are key topics for future work.
Grid losses were also not considered in this work, they depend on the specific conductors, the
current flowing and the length of the transmission line. However, this work has considered
distributed resources that, when compared with typical grid framework, reduce the required net
inflow from the grid, reduce grid current and hence the grid losses. Thus, in future works grid
losses should be considered. To account for actual grid capacities and losses, larger-scale
simulations based on the grid network models should be applied to enable an effective analysis
of grid stability.
Recent software applications like Google Now (Google Now, 2012) provides an mechanism
that is able to inference about the users habits and suggest what is best at the right time, the
technology proactively delivers to users information that it predicts (based on their search
habits) they may want. Thus, based on the consumer habits, a future approach should take in
consideration this type of tool, to advice, in real time the user about consumption measures that
can be taken in to account to increased efficiency and reduce energy costs.
BIBLIOGRAPHY
145
Bibliography
Aalami, H., Yousefi, G. R. & Parsa Moghadam, M. (2008). ‘Demand Response model
considering EDRP and TOU programs’. IEEE/PES Transmission and Distribution Conference
and Exposition, pp. 1–6. doi:10.1109/TDC.2008.4517059.
Ab, E., Hk, R. & Akkermans, H. (1996). ‘HOMEBOTS : Intelligent Agents for Decentralized
Load Management’, DA/DSM96 Europe Distribution Automation & Demand Side Management.
Pp. 1–12.
Abbass, H., What, A., Bender, A., Gaidow, S. & Whitbread, P. (2011). ‘Computational Red
Teaming: Past, Present and Future’. IEEE Computational Intelligence Magazine, Vol: 6, Issue:
1, pp. 30–42, doi: 10.1109/MCI.2010.939578.
Alessio, A. & Bemporad, A. (2007), ‘Decentralized Model Predictive Control of Constrained
Linear Systems’, in Proceedings of the European Control Conference, July 2007, EUCA, pp.
2813–2818.
Alessio, A., Barcelli, D. & Bemporad, A. (2011). ‘Decentralized model predictive control of
dynamically coupled linear systems’. Journal of Process Control, 21(5), pp. 705–714.
doi:10.1016/j.jprocont.2010.11.003.
Al-Hinai, A. & Feliachi, A., (2004).’ Application of Intelligent Control Agents in Power
Systems with Distributed Generators’. IEEE PES Power Systems Conference and Exposition,
pp: 1514 - 1519 Vol.3. doi:10.1109/PSCE.2004.1397709.
Allgöwer, F., Findeisen, R. & Nagy, Z. K. (2004). ‘Nonlinear Model Predictive Control: From
Theory to Application’, Journal of the Chinese Institute of Chemical Engineers vol. 35(3), pp.
299–315.
Anderson, B.D.O. & Moore, J.B.(1971). Linear Optimal Control, Prentice Hall 1971,
ISBN/ASIN: 0135368707.
Aung, H.N. Khambadkone, A.M., Srinivasan, D. & Logenthiran, T. (2008). ‘Agent-based
Intelligent Control for Real-time Operation of a Microgrid’. ICSET 2008. IEEE International
Conference on Sustainable Energy Technologies. doi: 10.1109/PEDES.2010.5712495
Badea, N., Epureanu, A., Ceanga, E. & Barbu, M. (2015). Design for Micro-Combined Cooling,
Heating and Power Systems. (N. Badea, Ed.). London: Springer London. doi:10.1007/978-1-
4471-6254-4.
Bălan, R., Stan, S. & Lăpușan, C. (2009). ‘A Model Based Predictive Control Algorithm for
BIBLIOGRAPHY
146
Building Temperature Control’. 3rd IEEE International Conference on Digital Ecosystems and
Technologies, pp. 540 – 545.
Balijepalli, V. S. K. M., Pradhan, V., Khaparde, S. A. & Shereef, R. M. (2011). ‘Review of
Demand Response under Smart Grid Paradigm’. IEEE PES Innovative Smart Grid
Technologies, pp. 236 – 243. ISBN: 978-1-4673-0316-3. doi: 10.1109/ISET-
India.2011.6145388.
Barata, F. A., Félix, N. & Neves-Silva, R. (2013b). ‘Distributed MPC for green thermally
comfortable buildings based on an electro-thermal modular approach’. Proceedings of
CETC2013 Conference on Electronics Telecommunications and Computers.
Barata, F. A., Félix, N. & Neves-Silva, R. (2014a). ‘Distributed MPC for green thermally
comfortable buildings based on an electro-thermal modular approach’. Procedia Technology
,Vol. 17, pp. 772–780.doi. doi:10.1016/j.protcy.2014.10.211.
Barata, F. A., Igreja, J. M. & Neves-Silva, R. (2012a). ‘Model Predictive Control for Thermal
House Comfort with Limited Energy Resources’, in Proceedings of the 10th Portuguese
Conference on Automatic Control, pp. 146-151.
Barata, F. A., Igreja, J. M. & Neves-Silva, R. (2014c). ‘Distributed MPC for Thermal Comfort
and Load Allocation with Energy Auction’, International Journal of Renewable Energy
Research (IJRER), Vol. 4 (2), pp. 371-383. Online ISSN: 1309-0127.
Barata, F. A. & Neves-Silva, R. (2012b). ‘Distributed Model Predictive Control for Thermal
House Comfort with Auction of Available Energy’, SG-TEP 2012: International Conference
On Smart Grid Technology, Economics And Policies, pp. 1-4. doi: 10.1109/SG-
TEP.2012.6642375.
Barata, F. A. & Neves-Silva, R. (2013a).’ Distributed Model Predictive Control for Housing
with Hourly Auction of Available Energy’, DoCEIS’13, 4th doctoral Conference on
Computing, Electrical and Industrial Systems, Technological Innovation for the Internet of
Things. IFIP Advances in Information and Communication Technology, Vol. 394, pp 469-476.
doi. 10.1007/978-3-642-37291-9_50. Online ISBN :978-3-642-37291-9.
Barata, F. A. & Neves-Silva, R. (2014b). ‘Distributed MPC for thermal comfort in buildings
with dynamically coupled zones and limited energy resources’, DoCEIS’14, Technological
Innovation for Collective Awareness Systems, IFIP Advances in Information and
Communication Technology, Vol. 423, pp. 305-312. doi: 10.1007/978-3-642-54734-8_34.
Online ISBN: 978-3-642-54734-8.
Barata, F., Campos, R. and Neves-Silva, R. (2013c). ‘Distributed MPC for Thermal House
BIBLIOGRAPHY
147
Comfort with Shifting Loads and Limited Energy Resources’. International Conference on
Renewable Energy Research and Applications (ICRERA), pp. 584 - 589, doi:
10.1109/ICRERA.2013.6749823.
Bemporad, A. (1997). Reference Governors: On-Line Set-Point Optimization Techniques for
Constraint Fulfillment. PhD Thesis. Universit`a di Firenze, Italy.
Bendtsen, J., Trangbaek, K. & Stoustrup, J. (2010). ‘Hierarchical model predictive control for
resource distribution’. 49th IEEE Conference on Decision and Control (CDC), pp. 2468–2473.
doi:10.1109/CDC.2010.5717038.
Bequette, B. (2003). Process Control, Modeling, Design and Simulation, Prentice Hall, pp.58.
Biegel, B., Stoustrup, J. (2014). ‘Distributed MPC Via Dual Decomposition’. Distributed Model
Predictive Control Made Easy Series: Intelligent Systems, Control and Automation: Science and
Engineering, Vol. 69 Chapter 11, pp 179-192.
Blanquet, A., Teixeira, C., Santos, J. & Alves, F. (2009). ‘From DA to SmartGrids – Evolution
or Revolution?’, (0770), 8–11. Print ISBN: 978-1-84919126-5.
Boom, T. J.J. & Stoorvogel A. A. (2010). Model Predictive Control. DISC Course, Lecture
Notes.
Borrelli, F., Keviczky, T. & Stewart, G.E. (2005), ‘Decentralized Constrained Optimal Control
Approach to Distributed Paper Machine Control’, in Proceedings of the 44th IEEE Conference
on Decision and Control and the European Control Conference, pp. 3037–3042.
Bsria, (2014), Available from: https://www.bsria.com/.
Callaway, D. S. (2009). ‘Tapping the energy storage potential in electric loads to deliver load
following and regulation, with application to wind energy’. Energy Conversion and
Management, 50(5), pp. 1389–1400. doi:10.1016/j.enconman.2008.12.012.
Camacho, E. F. (1993). ‘Constrained Generalized Predictive Control.’ IEEE Transactions on
Automatic Control, vol. 38(2), pp. 327-332.
Camponogara, E., Jia, D., Krogh, B.H. & Talukdar, S. (2002), ‘Distributed Model Predictive
Control’, IEEE Control Systems Magazine, 22, pp. 44–52.
Cecati, C., Mokryani, G., Piccolo, A. & Siano, P. (2010). ‘An overview on the smart grid
concept’. IECON 2010 - 36th Annual Conference on IEEE Industrial Electronics Society, pp.
3322–3327. doi:10.1109/IECON.2010.5675310.
Chebbo, M. (2007). ‘EU SmartGrids Framework Electricity Networks of the future 2020 and
beyond’, SmartGrids Technology Platform.
BIBLIOGRAPHY
148
Chen, C., Zhu, Y. & Xu, Y. (2010). ‘Distributed Generation and Demand Side Management’.
China International Conference on Electricity Distribution (CICED). Print ISBN: 978-1-4577-
0066-8.
Christofides, P. D., Scattolini, R., Muñoz de la Peña, D. & Liu, J. (2013). ‘Distributed model
predictive control: A tutorial review and future research directions’. Computers & Chemical
Engineering, pp. 51, 21–41. doi:10.1016/j.compchemeng.2012.05.011.
Clarke, D. W., Mohtadi, C. & Tuffs, P.S. (1987). ‘Generalized Predictive Control. Part I. The
Basic Algorithm’. Automatica, vol. 23, no 2, pp. 137-148.
Clarke, D. W., Mohtadi, C. & Tuffs, P.S. (1987b). ‘Generalized Predictive Control - Part II.
Extensions and Interpretations’, Automatica, vol. 23, no 2, pp. 149-160.
Commission, E. (2006), ‘Vision and Strategy for Europe’s Electricity Networks of the Future’,
European SmartGrids Technologic Platform. 2006.
Dimeas, A. L., Ieee, S. M., Hatziargyriou, N. D. & Member, S. (2005). ‘A MAS architecture for
Microgrids control’, Proceedings of the 13th International Conference on Intelligent Systems
Application to Power Systems, pp. 402–406. doi: 10.1109/ISAP.2005.1599297.
Doan, D., Keviczky, T., Necoara, I. and Diehl, M. (2009). ‘A jacobi algorithm for distributed
model predictive control of dynamically coupled systems’. American Control Conference, 2009.
DOE, U.S Department of Energy (2014). The Smart Grid: An Introduction. Available from:
http://energy.gov/sites/prod/files/oeprod/DocumentsandMedia/DOE_SG_Book_Single_Pages.p
df.
Du, X., Xi, Y. & Li, S. (2001), ‘Distributed Model Predictive Control for Large-scale Systems’,
in Proceedings of the American Control Conference, June 2001, pp. 3142–3143.
Dunbar, W.B. (2007), ‘Distributed Receding Horizon Control of Dynamically Coupled
Nonlinear Systems’, IEEE Transactions on Automatic Control, 52, 1249–1263.
Dunbar, W.B. & Murray, R.M. (2006), ‘Distributed Receding Horizon Control for Multi-
Vehicle Formation Stabilization’, Automatica, 42, pp. 549–558.
Dynamic Demand. Available from: http://www.dynamicdemand.co.uk. [November 2014].
EcoGrid (2014). Available from: http://www.eu-ecogrid.net/ [Dezember 2014].
EIA, U.S Energy Information Administration. Available from: http://www.eia.gov/ . [November
2014].
Fan, Z., Kulkarni, P., Gormus, S., Efthymiou, C., Kalogridis, G., Sooriyabandara, M., Chin, W.
BIBLIOGRAPHY
149
H. (2013). ‘Smart Grid Communications: Overview of Research Activities’, IEEE
Communications Surveys & Tutorials, Vol. 15(1), pp. 21 – 38. doi:
10.1109/SURV.2011.122211.00021.
Favre-Perrod, P., Critchley, R.; Catz, E.; Bazargan, M. (2009). ‘New participants in SmartGrids
and associated challenges in the transition towards the grid of the future’, Proceedings IEEE
PowerTech, pp: 1–5. doi: 10.1109/PTC.2009.5281828.
Ferber, J. (1999). Multi-Agent Systems: An Introduction to Distributed Artificial Intelligence.
Edinburgh: Addison Wesley, 1999. ISBN 0-201-36048-9.
Figueiredo, J., Martins, J. (2010). ‘Energy Production System Management - Renewable energy
power supply integration with Building Automation System’. Energy Conversion and
Management, 51(6), pp. 1120–1126. doi:10.1016/j.enconman.2009.12.020.
ForecastWacth (2014). Long-Term Analysis of Short-Term High Temperature Forecasts,
September 2006 through September 2014.
Franco, E., Parisini, T. & Polycarpou, M.M. (2007), ‘Design and Stability Analysis of
Cooperative Receding-horizon Control of Linear Discrete-time Agents’, International Journal
of Robust and Nonlinear Control, 17, pp. 982–1001.
Freire, R. Z., Oliveira, G. H. C. & Mendes, N. (2008). ‘Predictive controllers for thermal
comfort optimization and energy savings’. Journal Energy and Buildings, 40(7), pp. 1353–
1365. doi:10.1016/j.enbuild.2007.12.007.
Froisy, J. B. (1994). ‘Model predictive control: Past, present and future’. ISA Transactions ,
vol.33 ,pp. 235–243.
Fukushima, H. & Bitmead, R.R. (2005), ‘Robust Constrained Predictive Control using
Comparison Model’, Automatica, 41, pp. 97–106.
Giovanini, L., Balderud, J. & Katebi, M.R. (2007), ‘Autonomous and Decentralized Mission
Planning for Clusters of UUVs’, International Journal of Control, 80, pp. 1169–1179.
Google Now, (2012). Available from: https://www.google.com/landing/now/. [January 2015].
Greco, C., Menga, G., Mosca, E. & Zappa, G. (1984). “Performance improvements of self-
tuning controllers by multistep horizons: The MUSMAR approach,” Automatica, vol. 20, pp.
681–699.
Gulley, N., Chistol, G. (2006). Thermo Model of a House. Available from:
www.mathworks.com. [November 2011].
Hamidi, V. (2007). ‘New control methods in demand side management to improve the security
BIBLIOGRAPHY
150
of supply in the UK’s electricity network’. 42nd International Universities Power Engineering
Conference, pp. 132–137. doi:10.1109/UPEC.2007.4468933.
Hamidi, V., Li, F. & Robinson, F. (2008). ‘The effect of responsive demand in domestic sector
on power system operation in the networks with high penetration of renewables’. IEEE Power
and Energy Society General Meeting - Conversion and Delivery of Electrical Energy in the 21st
Century, pp. 1–8. doi:10.1109/PES.2008.4596269.
Hazyuk I, Ghiaus C & Penhouet D. (2012). ‘Optimal temperature control of intermittently
heated buildings using Model Predictive Control: Part I Building modelling’. Journal Building
and Environment 51 pp. 379-387. doi: 10.1016/j.buildenv.2011.11.009.
Hazyuk, I., Ghiaus, C. & Penhouet, D. (2012). ‘Optimal temperature control of intermittently
heated buildings using Model Predictive Control: Part II – Control algorithm’. Journal Building
and Environment, 51, pp. 388–394. doi:10.1016/j.buildenv.2011.11.008.
IEA, International Energy Agency (2011), ‘Smart Grids Technology Roadmap’. Available
from: https://www.iea.org/.../smartgrids_roadmap.pdf. [January 2015].
Igreja, J. M. & Cruces, J. C. (2002). ‘An I/0 Mixed Constrained Stabilizing MPC Applied to
Nonlinear Processes’. Proceedings of MED 2002: 10th Mediterranean Conference on Control
and Automation.
Igreja, J. M., Costa, S. J., Lemos, J. M. & Cadete, F. M. (2011). ‘Multi-Agent Predictive
Control with application in Intelligent Infrastructures’. Computational Intelligence and Decision
Making , 61: pp. 121-130 , doi:10.1007/978-94-007-4722-7_12. ISBN: 978-94-007-4721-0
James, G., Jones, T., (2010), ‘Modelling the ability of customer loads to track the output of
renewable generation’, 8th International Conference on Advances in Power System Control,
Operation and Management (APSCOM 2009). doi: 10.1049/cp.2009.1854.
Jia, D. & Krogh, B.H. (2002), ‘Min-Max Feedback Model Predictive Control for Distributed
Control with Communication’, in Proceedings of the American Control Conference, May 2002,
pp. 4507–4512.
Jian, Z., Qian, A., Chuanwen, J., Xingang, W., Zhanghua, Z. & Chenghong, G. (2009). ‘The
application of multi agent system in microgrid coordination control’, International Conference
on Sustainable Power Generation and Supply, pp. 1–6.
doi:10.1109/SUPERGEN.2009.5348277.
Jimenez, D. (2000). ‘Robustness aspects of Model Predictive Control’, PhD Thesis.
Jun, H. B (2009). The Development , Implementation , and Application of Demand Side
Management and control ( DSM + c ) Algorithm for Integrating Micro- generation System
BIBLIOGRAPHY
151
within Built Environment, PhD Thesis.
Keviczky, T., Borrelli, F. & Balas, G. J. (2006). ‘Decentralized receding horizon control for
large scale dynamically decoupled systems’. Automatica, 42(12), pp. 2105–2115.
doi:10.1016/j.automatica.2006.07.008.
Kim, T.H. & Sugie, T. (2005), ‘Robust Decentralized MPC Algorithm for a Class of
Dynamically Interconnected Constrained Systems’, in Proceedings of the IEEE Conference on
Decision and Control, pp. 290–295.
Koeppel, G., Korpas, M. (2006), ‘Using storage devices for compensating uncertainties caused
by non-dispatchable generators’, International Conference on Probabilistic Methods Applied to
Power Systems.doi: 10.1109/PMAPS.2006.360269. Print ISBN: 978-91-7178-585-5.
Korolija, I., Marjanovic-Halburd, L., Zhang, Y. & Hanby, V. I. (2011). ‘Influence of building
parameters and HVAC systems coupling on building energy performance’. Journal Energy and
Buildings, 43(6), pp. 1247–1253. doi:10.1016/j.enbuild.2011.01.003.
Kosek, A. M., Costanzo, G. T., Bindner, H. W. & Gehrke, O. (2013). ‘An overview of demand
side management control schemes for buildings in smart grids’, IEEE International Conference
on Smart Energy Grid Engineering (SEGE), pp. 1–9. doi:10.1109/SEGE.2013.6707934.
Krishnappa, K., (2008). Active Networks Demand Side Management & Voltage Control. Master
of Science in Energy Systems and the Environment.
Krogh, B. H. (2001). ‘Distributed model predictive control’. Proceedings of the 2001 American
Control Conference. (Cat. No.01CH37148), 5, pp. 2767–2772. doi:10.1109/ACC.2001.946306
Kuwata, Y., Richards, A., Schouwenaars, T. & How, J. (2007), ‘Distributed Robust Receding
Horizon Control for Multivehicle Guidance’, IEEE Transactions on Control Systems
Technology, 14, pp. 627–641.
Li, P., Song, B., Wang, W. & Wang, T. (2010). ‘Multi-agent approach for service restoration of
microgrid’. 5th IEEE Conference on Industrial Electronics and Applications, pp. 962–966.
doi:10.1109/ICIEA.2010.5515722.
Ling, K.V., Maciejowski, J.M. & Wu, B. (2005), ‘Multiplexed Model Predictive Control’, 16th
IFAC World Congress, pp. 574–579.
Löhnberg, P., Mulder, B. A. (1999). ‘Adaptative Simple Predictive Control for a High-
Efficiency Domestic Heather’. Journal A, Benelux Quarterly Journal on Automatic Control, vol
40, nº3, pp. 25-30.
Luo, T., Ault, G. & Galloway, S. (2010). ‘Demand Side Management in a Highly Decentralized
BIBLIOGRAPHY
152
Energy Future’. 45th International Universities Power Engineering Conference (UPEC). 2010.
Print ISBN: 978-1-4244-7667-1.
Ma, Y., Kelman, A., Daly, A. & Borrelli, F., (2012). ‘Predictive Control for Energy Efficient
Buildins with Thermal Storage’. IEEE Control System Magazine, vol 32, nº1, pp. 44 – 64.
Maciejowski, J.M. (2002), Predictive Control with Constraints, New York: Prentice-Hall.
Magni, L. & Scattolini, R. (2006), ‘Stabilizing Decentralized Model Predictive Control of
Nonlinear Systems’, Automatica, 42, pp. 1231–1236.
Mahmood, A., Ullah, M. N., Razzaq, S., Basit, A., Mustafa, U., Naeem, M. & Javaid, N. (2014)
‘A New Scheme for Demand Side Management in Future Smart Grid Networks’, Procedia
Computer Science, vol.32, pp. 477–484, doi: 10.1016/j.procs.2014.05.450.
Maile, T., Fischer, M., Bazjanac, V. (2007). Building Energy Performance Simulation Tools - a
Life-Cycle and Interoperable Perspective. Stanford University.
Mayne, D.Q., Rawlings, J.B., Rao, C.V. & Scokaert, P.O.M. (2000), ‘Constrained Model
Predictive Control: Stability and Optimality’, Automatica, 36, pp. 789–814.
Mayne, D.Q., Seron, M.M. & Rakovic´ , S.V. (2005), ‘Robust Model Predictive Control of
Constrained Linear Systems with Bounded Disturbances’, Automatica, 41, 219–224.
Mendes, N., Oliveira G. & Araújo, H. (2001). ‘Building Thermal Performance Analysis by
Using Matlab/Simulink.’ Proceedings of the 7th International IBPSA Conference, pp. 473-480.
Molderink, A., Member, S., Bakker, V., Bosman, M. G. C., Hurink, J. L. & Smit, G. J. M.
(2010). ‘Management and Control of Domestic Smart Grid Technology’, IEEE Transactions on
Smart Grid, Volume: 1, Issue: 2, pp. 109 – 119.
Morosan, P.-D. Bourdais, R.; Dumur, D.; Buisson, J. (2010). ‘A distributed MPC applied to
multisource temperature regulation in buildings’, Proceedings of the 2nd IFAC Workshop on
Distributed Estimation and Control in Networked Systems, pp. 91–96.
Mosca, E. (1995). Optimal Predictive and Adaptative Control. Prentice Hall. ISBN 0-13-
847609-8.
Mosca, E., Lemos, J. M., Zhang, J. (1990). ‘Stabilizing I/O receding horizon control’,
Proceedings of the 29th IEEE Conference on Decision and Control, vol. 4, pp. 2518 – 2523.doi:
10.1109/CDC.1990.203454.
Mosca, E., Zappa, G. and Lemos, J. M. (1989). “Robustness of multipredictor adaptive
regulators: MUSMAR”. Automatica, vol. 25(4), pp. 521–529.
Müller, M. A., Reble, M. and Allgöwer, F. (2012), ‘Cooperative control of dynamically
BIBLIOGRAPHY
153
decoupled systems via distributed model predictive control’. Int. J. Robust Nonlinear Control,
22,pp. 1376–1397. doi: 10.1002/rnc.2826.
NREL (2014). National Renewable Energy Laboratory. Solar and Wind Forecasting. Available
from: http://www.nrel.gov/electricity/transmission/resource_forecasting.html. [December
2014].
Navetas Energy Management. Available from: http://www.navetas.com. [November 2014].
Negenborn, R. (2007.). Multi-Agent Model Predictive Control with Applications to Power
Networks. PhD Thesis. ISBN 9789055840939.
Negenborn, R. R., Schutter, B. De, Hellendoorn, H. (2006). ‘Multi-agent model predictive
control for transportation networks with continuous and discrete elements’. 11th IFAC
Symposium on Control in Transportation Systems, pp. 609-614. doi: 10.3182/20060829-3-NL-
2908.00105.
Nolte, I. & Strong, D. (2011). Europe’s buildings under the microscope. Buildings Performance
Institute Europe, 2011. ISBN: 9789491143014
Nunes, C. S., Mendonça, T., Lemos, J. M. & Amorim, P. (2007). ‘Control of depth of anesthesia
using MUSMAR - Exploring electromyography and the analgesic dose as accessible
disturbances’. Proceedings of Annual International Conference of the IEEE Engineering in
Medicine and Biology , pp. 1574–1577. doi:10.1109/IEMBS.2007.4352605.
Orukpe, P. E. (2005). Basics of Model Predictive Control. Available from:
http://www3.imperial.ac.uk/portal/pls/portallive/docs/1/50918.PDF
Paul,S., Rabbani, M. S., Kundu, R. K. & Zaman, S. M. R. (2014). ‘A review of smart
technology (Smart Grid) and its features’, 1st Int. Conf. Non Conv. Energy (ICONCE 2014), pp.
200–203, doi: 10.1109/ICONCE.2014.6808719
Pipattanasomporn, M., Feroze, H., Rahman, S. (2009). ’Multi-agent systems in a distributed
smart grid: Design and implementation’. IEEE/PES Power Systems Conference and Exposition,
pp. 1–8. doi:10.1109/PSCE.2009.4840087.
Qin, S., Badgwell, T. (2003). ‘A survey of industrial model predictive control technology’.
Control Engineering Practice, vol. 11(7), pp. 733–764.
Raffard, R.L., Tomlin, C.J. & Boyd, S.P. (2004), ‘Distributed Optimization for Cooperative
Agents: Application to Formation Flight’, in Proceedings of the 43rd IEEE Conference on
Decision and Control, pp. 2453–2459.
Rato, L., Borrelli, D.; Mosca, E.; Lemos, J.M. ; Balsa, P. (1997). ‘MUSMAR based switching
BIBLIOGRAPHY
154
control of a solar collector field’. ECC97, Proceedings of European Control Conference.
Raza, S.M.A. Akbar, M.; Kamran, F. (2005). ‘Use Case Model of Genetic Algorithms of Agents
for Control of Distributed Power System Networks’, Proceedings of the IEEE Symposium on
Emerging Technologies, pp. 405–411. doi: 10.1109/ICET.2005.1558916.
Ribeiro, P. F., Johnson B. K., Crow, M. L., Arsoy, A. & Liu Y. (2001), ‘Energy storage systems
for advanced power applications’, Proceedings of the IEEE, 89(12), pp. 1744–1756.
Richards, A. & How, J.P. (2007), ‘Robust Distributed Model Predictive Control’, International
Journal of Control, 80, pp. 1517–1531.
Riggs, D. J. & Bitmead, R. R. (2010). ‘Negotiation of coupled constraints in coordinated
vehicles’. 49th IEEE Conference on Decision and Control (CDC), pp. 479–484.
doi:10.1109/CDC.2010.5717221.
Roche, R., Blunier, B., Miraoui, A., Hilaire, V. & Koukam, A. (2010). ‘Multi-agent systems for
grid energy management: A short review’. IECON 2010 - 36th Annual Conference on IEEE
Industrial Electronics Society, pp. 3341–3346. doi:10.1109/IECON.2010.5675295.
Rugh, W. J. (1996). Linear System Theory. Prentice Hall 1996. ISBN 0-13-441205-2.
Saffre, F. & Gedge, R. (2010). ‘Demand-Side Management for the Smart Grid’. IEEE/IFIP
Network Operations and Management Symposium Workshops, pp. 300–303.
doi:10.1109/NOMSW.2010.5486558.
Saluje, K. K., Ramanathan, P., Gokce, E. I., Varshney, P. K., Kassam, A., Poor, H. V, Weyman,
J. (2011). ‘Distributed Optimization for Model Predictive Control of Linear Dynamic Networks
With Control-Input and Output Constraints’, IEEE Transactions on Automation Science and
Engineering, pp. 2338(1), 233–242.
Savage, H., Kennedy, J., Fox, B. & Flynn, D. (2008). ‘Managing variability of wind energy
with heating load control’. 43rd International Universities Power Engineering Conference, pp.
1–5. doi:10.1109/UPEC.2008.4651571.
Scattolini, R. (2009). ‘Architectures for distributed and hierarchical Model Predictive Control –
A review’. Journal of Process Control, 19(5), pp. 723–731. doi:10.1016/j.jprocont.2009.02.003
Scokaert, P.O.M. & Mayne, D.Q. (1998), ‘MinMax Feedback Model Predictive Control for
Constrained Linear Systems’, IEEE Transactions on Automatic Control, 43, pp. 1136–1142.
Shaikh, S.K.M. Dharme, A.A. (2009). ‘Time of Use P Pricing – India , a Case Study’, ICPS '09
International Conference on Power Systems. doi: 10.1109/ICPWS.2009.5442760.
Siano, P. (2014). ‘Demand response and smart grids - A survey’, Renewable. Sustainable
BIBLIOGRAPHY
155
Energy Reviews, vol. 30, pp. 461–478. doi:10.1016/j.rser.2013.10.022.
Siemens (2014), Communications network solutions for smart grids, Available on:
http://w3.siemens.com/smartgrid/global/drafts/services/metering-communications-
services/smart-communications-solutions/Documents/Communications-network-solutions-for-
Smart-Grid.pdf [September 2014].
Stewart, B. T., Rawlings, J. B. & Wright, S. J. (2010). ‘Hierarchical cooperative distributed
model predictive control’, American Control Conference (ACC), pp. 3963 - 3968 ISSN 0743-
1619.
StorePET Project, Available from: http://www.storepet-fp7.eu/project-overview. [November
2014].
Thavlov, A. (2008). Dynamic Optimization of Power Consumption, Master Thesis.
Trodden, P. & Richards, A. (2010). ‘Distributed model predictive control of linear systems with
persistent disturbances’. International Journal of Control, 83(8), pp. 1653–1663.
doi:10.1080/00207179.2010.485280.
Trodden, P.A. (2009), Robust Distributed Control of Constrained Linear Systems, Ph.D.
dissertation, University of Bristol.
Venkat, A. N.(2006). Distributed Model Predictive Control: Theory and Applications, Phd
Thesis 2006.
Venkat, A. N., Hiskens, I. a., Rawlings, J. B. & Wright, S. J. (2006). ‘Distributed Output
Feedback MPC for Power System Control’. Proceedings of the 45th IEEE Conference on
Decision and Control, pp. 4038–4045. doi:10.1109/CDC.2006.377176.
Venkat, A.N., Hiskens, I.A., Rawlings, J.B. & Wright, S.J. (2008), ‘Distributed MPC Strategies
with Application to Power System Automatic Generation Control’, IEEE Transactions on
Control Systems Technology, 16, pp. 1192–1206.
Venkat, A.N., Rawlings, J.B. & Wright, S.J. (2004), ‘Plant-wide Optimal Control with
Decentralized MPC’, Proceedings of the 7th IFAC International Symposium on Dynamics and
Control of Process Systems.
Virote, J. & Neves-Silva, R. (2012). ‘Stochastic models for building energy prediction based on
occupant behavior assessment’, Energy and Buildings., vol. 53, pp. 183–193, doi:
10.1016/j.enbuild.2012.06.001.
VTT Thecnology (2014). Wind power forecasting accuracy and uncertainty in Finland.
Wang, C. & Ong, C.-J. (2010). ‘Distributed model predictive control of dynamically decoupled
BIBLIOGRAPHY
156
systems with coupled cost’. Automatica, 46(12), pp. 2053–2058.
doi:10.1016/j.automatica.2010.09.002.
Wang, C., Zhou, Y., Jiao, B., Wang, Y., Liu, W. & Wang, D. (2015). ‘Robust optimization for
load scheduling of a smart home with photovoltaic system’. Energy Conversion and
Management, pp. 1–11. doi:10.1016/j.enconman.2015.01.053
Waslander, S.L., Inalhan, G. & Tomlin, C.J. (2004), ‘Decentralized Optimization via Nash
Bargaining’, in Theory and Algorithms for Cooperative Systems, Series on Computers and
Operations Research (Vol. 4, Chapter 25), eds. D. Grundel, R. Murphey & P.M. Pardolos,
World Scientific, Singapore, pp. 565–582.
Wedde, H. F., Lehnhoff, S., Handschin, E. & Krause, O. (2006). ‘Real-Time Multi-Agent
Support for Decentralized Management of Electric Power’. 18th Euromicro Conference on
Real-Time Systems (ECRTS’06), pp. 43–51. doi:10.1109/ECRTS.2006.22.
Wedde, H. F., Lehnhoff, S., Moritz, K. M., Handschin, E. & Krause, O. (2008). ‘Distributed
Learning Strategies for Collaborative Agents in Adaptive Decentralized Power Systems’. 15th
Annual IEEE International Conference and Workshop on the Engineering of Computer Based
Systems, pp. 26–35. doi:10.1109/ECBS.2008.59.
Werbo, P. (2011). ‘Computational Intelligence for the Smart Grid-History, Challenges, and
Opportunities’. IEEE Computational Intelligence Magazine, vol. 6, no. 3, pp. 14-21, 2011.
doi.: 10.1109/MCI.2011.941587.
Xu, Q., Jia, X. & He L., (2010). ‘The control of Distributed Generation System using Multi-
Agent System’, International Conference On Electronics and Information Engineering (ICEIE),
Vol 1.pp 30 – 33, doi: 10.1109/ICEIE.2010.5559832.
Yang, H., Zhang, Y. & Tong, X. (2006). ‘System Dynamics Model for Demand Side
Management’. 3rd International Conference on Electrical and Electronics Engineering, pp. 1–
4. doi:10.1109/ICEEE.2006.251854.
Yu, N., Yu, J. (2006). ‘Optimal TOU Decision Considering Demand Response Model’,
PowerCon. International Conference on Power System Technology, pp. 1-6. doi:
10.1109/ICPST.2006.321461.
Zaidi, A. A., Zia, T. & Kupzog, F. (2010). ‘Automated demand side management in microgrids
using load recognition’. 8th IEEE International Conference on Industrial Informatics, pp. 774–
779. doi:10.1109/INDIN.2010.5549646.
Zeng, J., Liu, J. F., Ngan, H. W. & Wu, J. (2009). ‘A multi-agent solution to energy
management of distributed hybrid renewable energy generated system’, 8th International
BIBLIOGRAPHY
157
Conference on Advances in Power System Control, Operation and Management (APSCOM
2009), pp. 1 – 6.
Zeng, T. (2010), Model Predictive Control. Scyco. ISBN 978-953-307-102-2.
Zhang, Y. & Li, S. (2007). ‘Networked model predictive control based on neighbourhood
optimization for serially connected large-scale processes’. Journal of Process Control, 17(1),
pp.37–50. doi:10.1016/j.jprocont.2006.08.009.
Zong, Y., Kullmann, D., Thavlov, A., Gehrke, O. & Bindner H. W. (2012). ‘Application of
predictive control for active load management in a distributed power system with high wind
penetration’, IEEE Transactions on Smart Grid, vol. 3 Nº 2, pp 1055-1062.
158